1 // SPDX-License-Identifier: GPL-2.0
3 * Shared application/kernel submission and completion ring pairs, for
4 * supporting fast/efficient IO.
6 * A note on the read/write ordering memory barriers that are matched between
7 * the application and kernel side.
9 * After the application reads the CQ ring tail, it must use an
10 * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses
11 * before writing the tail (using smp_load_acquire to read the tail will
12 * do). It also needs a smp_mb() before updating CQ head (ordering the
13 * entry load(s) with the head store), pairing with an implicit barrier
14 * through a control-dependency in io_get_cqring (smp_store_release to
15 * store head will do). Failure to do so could lead to reading invalid
18 * Likewise, the application must use an appropriate smp_wmb() before
19 * writing the SQ tail (ordering SQ entry stores with the tail store),
20 * which pairs with smp_load_acquire in io_get_sqring (smp_store_release
21 * to store the tail will do). And it needs a barrier ordering the SQ
22 * head load before writing new SQ entries (smp_load_acquire to read
25 * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application
26 * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after*
27 * updating the SQ tail; a full memory barrier smp_mb() is needed
30 * Also see the examples in the liburing library:
32 * git://git.kernel.dk/liburing
34 * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens
35 * from data shared between the kernel and application. This is done both
36 * for ordering purposes, but also to ensure that once a value is loaded from
37 * data that the application could potentially modify, it remains stable.
39 * Copyright (C) 2018-2019 Jens Axboe
40 * Copyright (c) 2018-2019 Christoph Hellwig
42 #include <linux/kernel.h>
43 #include <linux/init.h>
44 #include <linux/errno.h>
45 #include <linux/syscalls.h>
46 #include <linux/compat.h>
47 #include <net/compat.h>
48 #include <linux/refcount.h>
49 #include <linux/uio.h>
50 #include <linux/bits.h>
52 #include <linux/sched/signal.h>
54 #include <linux/file.h>
55 #include <linux/fdtable.h>
57 #include <linux/mman.h>
58 #include <linux/percpu.h>
59 #include <linux/slab.h>
60 #include <linux/blkdev.h>
61 #include <linux/bvec.h>
62 #include <linux/net.h>
64 #include <net/af_unix.h>
66 #include <linux/anon_inodes.h>
67 #include <linux/sched/mm.h>
68 #include <linux/uaccess.h>
69 #include <linux/nospec.h>
70 #include <linux/sizes.h>
71 #include <linux/hugetlb.h>
72 #include <linux/highmem.h>
73 #include <linux/namei.h>
74 #include <linux/fsnotify.h>
75 #include <linux/fadvise.h>
76 #include <linux/eventpoll.h>
77 #include <linux/splice.h>
78 #include <linux/task_work.h>
79 #include <linux/pagemap.h>
80 #include <linux/io_uring.h>
82 #define CREATE_TRACE_POINTS
83 #include <trace/events/io_uring.h>
85 #include <uapi/linux/io_uring.h>
90 #define IORING_MAX_ENTRIES 32768
91 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
94 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
96 #define IORING_FILE_TABLE_SHIFT 9
97 #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
98 #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
99 #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
100 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
101 IORING_REGISTER_LAST + IORING_OP_LAST)
103 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
104 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
108 u32 head ____cacheline_aligned_in_smp;
109 u32 tail ____cacheline_aligned_in_smp;
113 * This data is shared with the application through the mmap at offsets
114 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
116 * The offsets to the member fields are published through struct
117 * io_sqring_offsets when calling io_uring_setup.
121 * Head and tail offsets into the ring; the offsets need to be
122 * masked to get valid indices.
124 * The kernel controls head of the sq ring and the tail of the cq ring,
125 * and the application controls tail of the sq ring and the head of the
128 struct io_uring sq, cq;
130 * Bitmasks to apply to head and tail offsets (constant, equals
133 u32 sq_ring_mask, cq_ring_mask;
134 /* Ring sizes (constant, power of 2) */
135 u32 sq_ring_entries, cq_ring_entries;
137 * Number of invalid entries dropped by the kernel due to
138 * invalid index stored in array
140 * Written by the kernel, shouldn't be modified by the
141 * application (i.e. get number of "new events" by comparing to
144 * After a new SQ head value was read by the application this
145 * counter includes all submissions that were dropped reaching
146 * the new SQ head (and possibly more).
152 * Written by the kernel, shouldn't be modified by the
155 * The application needs a full memory barrier before checking
156 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
162 * Written by the application, shouldn't be modified by the
167 * Number of completion events lost because the queue was full;
168 * this should be avoided by the application by making sure
169 * there are not more requests pending than there is space in
170 * the completion queue.
172 * Written by the kernel, shouldn't be modified by the
173 * application (i.e. get number of "new events" by comparing to
176 * As completion events come in out of order this counter is not
177 * ordered with any other data.
181 * Ring buffer of completion events.
183 * The kernel writes completion events fresh every time they are
184 * produced, so the application is allowed to modify pending
187 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
190 enum io_uring_cmd_flags {
191 IO_URING_F_NONBLOCK = 1,
192 IO_URING_F_COMPLETE_DEFER = 2,
195 struct io_mapped_ubuf {
198 struct bio_vec *bvec;
199 unsigned int nr_bvecs;
200 unsigned long acct_pages;
206 struct list_head list;
213 struct fixed_rsrc_table {
217 struct fixed_rsrc_ref_node {
218 struct percpu_ref refs;
219 struct list_head node;
220 struct list_head rsrc_list;
221 struct fixed_rsrc_data *rsrc_data;
222 void (*rsrc_put)(struct io_ring_ctx *ctx,
223 struct io_rsrc_put *prsrc);
224 struct llist_node llist;
228 struct fixed_rsrc_data {
229 struct fixed_rsrc_table *table;
230 struct io_ring_ctx *ctx;
232 struct fixed_rsrc_ref_node *node;
233 struct percpu_ref refs;
234 struct completion done;
239 struct list_head list;
245 struct io_restriction {
246 DECLARE_BITMAP(register_op, IORING_REGISTER_LAST);
247 DECLARE_BITMAP(sqe_op, IORING_OP_LAST);
248 u8 sqe_flags_allowed;
249 u8 sqe_flags_required;
254 IO_SQ_THREAD_SHOULD_STOP = 0,
255 IO_SQ_THREAD_SHOULD_PARK,
260 atomic_t park_pending;
263 /* ctx's that are using this sqd */
264 struct list_head ctx_list;
266 struct task_struct *thread;
267 struct wait_queue_head wait;
269 unsigned sq_thread_idle;
275 struct completion exited;
276 struct callback_head *park_task_work;
279 #define IO_IOPOLL_BATCH 8
280 #define IO_COMPL_BATCH 32
281 #define IO_REQ_CACHE_SIZE 32
282 #define IO_REQ_ALLOC_BATCH 8
284 struct io_comp_state {
285 struct io_kiocb *reqs[IO_COMPL_BATCH];
287 unsigned int locked_free_nr;
288 /* inline/task_work completion list, under ->uring_lock */
289 struct list_head free_list;
290 /* IRQ completion list, under ->completion_lock */
291 struct list_head locked_free_list;
294 struct io_submit_link {
295 struct io_kiocb *head;
296 struct io_kiocb *last;
299 struct io_submit_state {
300 struct blk_plug plug;
301 struct io_submit_link link;
304 * io_kiocb alloc cache
306 void *reqs[IO_REQ_CACHE_SIZE];
307 unsigned int free_reqs;
312 * Batch completion logic
314 struct io_comp_state comp;
317 * File reference cache
321 unsigned int file_refs;
322 unsigned int ios_left;
327 struct percpu_ref refs;
328 } ____cacheline_aligned_in_smp;
332 unsigned int compat: 1;
333 unsigned int cq_overflow_flushed: 1;
334 unsigned int drain_next: 1;
335 unsigned int eventfd_async: 1;
336 unsigned int restricted: 1;
339 * Ring buffer of indices into array of io_uring_sqe, which is
340 * mmapped by the application using the IORING_OFF_SQES offset.
342 * This indirection could e.g. be used to assign fixed
343 * io_uring_sqe entries to operations and only submit them to
344 * the queue when needed.
346 * The kernel modifies neither the indices array nor the entries
350 unsigned cached_sq_head;
353 unsigned sq_thread_idle;
354 unsigned cached_sq_dropped;
355 unsigned cached_cq_overflow;
356 unsigned long sq_check_overflow;
358 /* hashed buffered write serialization */
359 struct io_wq_hash *hash_map;
361 struct list_head defer_list;
362 struct list_head timeout_list;
363 struct list_head cq_overflow_list;
365 struct io_uring_sqe *sq_sqes;
366 } ____cacheline_aligned_in_smp;
369 struct mutex uring_lock;
370 wait_queue_head_t wait;
371 } ____cacheline_aligned_in_smp;
373 struct io_submit_state submit_state;
375 struct io_rings *rings;
377 /* Only used for accounting purposes */
378 struct mm_struct *mm_account;
380 const struct cred *sq_creds; /* cred used for __io_sq_thread() */
381 struct io_sq_data *sq_data; /* if using sq thread polling */
383 struct wait_queue_head sqo_sq_wait;
384 struct list_head sqd_list;
387 * If used, fixed file set. Writers must ensure that ->refs is dead,
388 * readers must ensure that ->refs is alive as long as the file* is
389 * used. Only updated through io_uring_register(2).
391 struct fixed_rsrc_data *file_data;
392 unsigned nr_user_files;
394 /* if used, fixed mapped user buffers */
395 unsigned nr_user_bufs;
396 struct io_mapped_ubuf *user_bufs;
398 struct user_struct *user;
400 struct completion ref_comp;
402 #if defined(CONFIG_UNIX)
403 struct socket *ring_sock;
406 struct xarray io_buffers;
408 struct xarray personalities;
412 unsigned cached_cq_tail;
415 atomic_t cq_timeouts;
416 unsigned cq_last_tm_flush;
417 unsigned long cq_check_overflow;
418 struct wait_queue_head cq_wait;
419 struct fasync_struct *cq_fasync;
420 struct eventfd_ctx *cq_ev_fd;
421 } ____cacheline_aligned_in_smp;
424 spinlock_t completion_lock;
427 * ->iopoll_list is protected by the ctx->uring_lock for
428 * io_uring instances that don't use IORING_SETUP_SQPOLL.
429 * For SQPOLL, only the single threaded io_sq_thread() will
430 * manipulate the list, hence no extra locking is needed there.
432 struct list_head iopoll_list;
433 struct hlist_head *cancel_hash;
434 unsigned cancel_hash_bits;
435 bool poll_multi_file;
437 spinlock_t inflight_lock;
438 struct list_head inflight_list;
439 } ____cacheline_aligned_in_smp;
441 struct delayed_work rsrc_put_work;
442 struct llist_head rsrc_put_llist;
443 struct list_head rsrc_ref_list;
444 spinlock_t rsrc_ref_lock;
446 struct io_restriction restrictions;
449 struct callback_head *exit_task_work;
451 struct wait_queue_head hash_wait;
453 /* Keep this last, we don't need it for the fast path */
454 struct work_struct exit_work;
455 struct list_head tctx_list;
458 struct io_uring_task {
459 /* submission side */
461 struct wait_queue_head wait;
462 const struct io_ring_ctx *last;
464 struct percpu_counter inflight;
468 spinlock_t task_lock;
469 struct io_wq_work_list task_list;
470 unsigned long task_state;
471 struct callback_head task_work;
475 * First field must be the file pointer in all the
476 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
478 struct io_poll_iocb {
480 struct wait_queue_head *head;
484 struct wait_queue_entry wait;
487 struct io_poll_remove {
497 struct io_timeout_data {
498 struct io_kiocb *req;
499 struct hrtimer timer;
500 struct timespec64 ts;
501 enum hrtimer_mode mode;
506 struct sockaddr __user *addr;
507 int __user *addr_len;
509 unsigned long nofile;
529 struct list_head list;
530 /* head of the link, used by linked timeouts only */
531 struct io_kiocb *head;
534 struct io_timeout_rem {
539 struct timespec64 ts;
544 /* NOTE: kiocb has the file as the first member, so don't do it here */
552 struct sockaddr __user *addr;
559 struct user_msghdr __user *umsg;
565 struct io_buffer *kbuf;
571 struct filename *filename;
573 unsigned long nofile;
576 struct io_rsrc_update {
602 struct epoll_event event;
606 struct file *file_out;
607 struct file *file_in;
614 struct io_provide_buf {
628 const char __user *filename;
629 struct statx __user *buffer;
641 struct filename *oldpath;
642 struct filename *newpath;
650 struct filename *filename;
653 struct io_completion {
655 struct list_head list;
659 struct io_async_connect {
660 struct sockaddr_storage address;
663 struct io_async_msghdr {
664 struct iovec fast_iov[UIO_FASTIOV];
665 /* points to an allocated iov, if NULL we use fast_iov instead */
666 struct iovec *free_iov;
667 struct sockaddr __user *uaddr;
669 struct sockaddr_storage addr;
673 struct iovec fast_iov[UIO_FASTIOV];
674 const struct iovec *free_iovec;
675 struct iov_iter iter;
677 struct wait_page_queue wpq;
681 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
682 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
683 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
684 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
685 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
686 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
692 REQ_F_LINK_TIMEOUT_BIT,
694 REQ_F_NEED_CLEANUP_BIT,
696 REQ_F_BUFFER_SELECTED_BIT,
697 REQ_F_NO_FILE_TABLE_BIT,
698 REQ_F_LTIMEOUT_ACTIVE_BIT,
699 REQ_F_COMPLETE_INLINE_BIT,
702 /* not a real bit, just to check we're not overflowing the space */
708 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
709 /* drain existing IO first */
710 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
712 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
713 /* doesn't sever on completion < 0 */
714 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
716 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
717 /* IOSQE_BUFFER_SELECT */
718 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
720 /* fail rest of links */
721 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
722 /* on inflight list, should be cancelled and waited on exit reliably */
723 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
724 /* read/write uses file position */
725 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
726 /* must not punt to workers */
727 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
728 /* has or had linked timeout */
729 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
731 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
733 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
734 /* already went through poll handler */
735 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
736 /* buffer already selected */
737 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
738 /* doesn't need file table for this request */
739 REQ_F_NO_FILE_TABLE = BIT(REQ_F_NO_FILE_TABLE_BIT),
740 /* linked timeout is active, i.e. prepared by link's head */
741 REQ_F_LTIMEOUT_ACTIVE = BIT(REQ_F_LTIMEOUT_ACTIVE_BIT),
742 /* completion is deferred through io_comp_state */
743 REQ_F_COMPLETE_INLINE = BIT(REQ_F_COMPLETE_INLINE_BIT),
744 /* caller should reissue async */
745 REQ_F_REISSUE = BIT(REQ_F_REISSUE_BIT),
749 struct io_poll_iocb poll;
750 struct io_poll_iocb *double_poll;
753 struct io_task_work {
754 struct io_wq_work_node node;
755 task_work_func_t func;
759 * NOTE! Each of the iocb union members has the file pointer
760 * as the first entry in their struct definition. So you can
761 * access the file pointer through any of the sub-structs,
762 * or directly as just 'ki_filp' in this struct.
768 struct io_poll_iocb poll;
769 struct io_poll_remove poll_remove;
770 struct io_accept accept;
772 struct io_cancel cancel;
773 struct io_timeout timeout;
774 struct io_timeout_rem timeout_rem;
775 struct io_connect connect;
776 struct io_sr_msg sr_msg;
778 struct io_close close;
779 struct io_rsrc_update rsrc_update;
780 struct io_fadvise fadvise;
781 struct io_madvise madvise;
782 struct io_epoll epoll;
783 struct io_splice splice;
784 struct io_provide_buf pbuf;
785 struct io_statx statx;
786 struct io_shutdown shutdown;
787 struct io_rename rename;
788 struct io_unlink unlink;
789 /* use only after cleaning per-op data, see io_clean_op() */
790 struct io_completion compl;
793 /* opcode allocated if it needs to store data for async defer */
796 /* polled IO has completed */
802 struct io_ring_ctx *ctx;
805 struct task_struct *task;
808 struct io_kiocb *link;
809 struct percpu_ref *fixed_rsrc_refs;
812 * 1. used with ctx->iopoll_list with reads/writes
813 * 2. to track reqs with ->files (see io_op_def::file_table)
815 struct list_head inflight_entry;
817 struct io_task_work io_task_work;
818 struct callback_head task_work;
820 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
821 struct hlist_node hash_node;
822 struct async_poll *apoll;
823 struct io_wq_work work;
826 struct io_tctx_node {
827 struct list_head ctx_node;
828 struct task_struct *task;
829 struct io_ring_ctx *ctx;
832 struct io_defer_entry {
833 struct list_head list;
834 struct io_kiocb *req;
839 /* needs req->file assigned */
840 unsigned needs_file : 1;
841 /* hash wq insertion if file is a regular file */
842 unsigned hash_reg_file : 1;
843 /* unbound wq insertion if file is a non-regular file */
844 unsigned unbound_nonreg_file : 1;
845 /* opcode is not supported by this kernel */
846 unsigned not_supported : 1;
847 /* set if opcode supports polled "wait" */
849 unsigned pollout : 1;
850 /* op supports buffer selection */
851 unsigned buffer_select : 1;
852 /* must always have async data allocated */
853 unsigned needs_async_data : 1;
854 /* should block plug */
856 /* size of async data needed, if any */
857 unsigned short async_size;
860 static const struct io_op_def io_op_defs[] = {
861 [IORING_OP_NOP] = {},
862 [IORING_OP_READV] = {
864 .unbound_nonreg_file = 1,
867 .needs_async_data = 1,
869 .async_size = sizeof(struct io_async_rw),
871 [IORING_OP_WRITEV] = {
874 .unbound_nonreg_file = 1,
876 .needs_async_data = 1,
878 .async_size = sizeof(struct io_async_rw),
880 [IORING_OP_FSYNC] = {
883 [IORING_OP_READ_FIXED] = {
885 .unbound_nonreg_file = 1,
888 .async_size = sizeof(struct io_async_rw),
890 [IORING_OP_WRITE_FIXED] = {
893 .unbound_nonreg_file = 1,
896 .async_size = sizeof(struct io_async_rw),
898 [IORING_OP_POLL_ADD] = {
900 .unbound_nonreg_file = 1,
902 [IORING_OP_POLL_REMOVE] = {},
903 [IORING_OP_SYNC_FILE_RANGE] = {
906 [IORING_OP_SENDMSG] = {
908 .unbound_nonreg_file = 1,
910 .needs_async_data = 1,
911 .async_size = sizeof(struct io_async_msghdr),
913 [IORING_OP_RECVMSG] = {
915 .unbound_nonreg_file = 1,
918 .needs_async_data = 1,
919 .async_size = sizeof(struct io_async_msghdr),
921 [IORING_OP_TIMEOUT] = {
922 .needs_async_data = 1,
923 .async_size = sizeof(struct io_timeout_data),
925 [IORING_OP_TIMEOUT_REMOVE] = {
926 /* used by timeout updates' prep() */
928 [IORING_OP_ACCEPT] = {
930 .unbound_nonreg_file = 1,
933 [IORING_OP_ASYNC_CANCEL] = {},
934 [IORING_OP_LINK_TIMEOUT] = {
935 .needs_async_data = 1,
936 .async_size = sizeof(struct io_timeout_data),
938 [IORING_OP_CONNECT] = {
940 .unbound_nonreg_file = 1,
942 .needs_async_data = 1,
943 .async_size = sizeof(struct io_async_connect),
945 [IORING_OP_FALLOCATE] = {
948 [IORING_OP_OPENAT] = {},
949 [IORING_OP_CLOSE] = {},
950 [IORING_OP_FILES_UPDATE] = {},
951 [IORING_OP_STATX] = {},
954 .unbound_nonreg_file = 1,
958 .async_size = sizeof(struct io_async_rw),
960 [IORING_OP_WRITE] = {
962 .unbound_nonreg_file = 1,
965 .async_size = sizeof(struct io_async_rw),
967 [IORING_OP_FADVISE] = {
970 [IORING_OP_MADVISE] = {},
973 .unbound_nonreg_file = 1,
978 .unbound_nonreg_file = 1,
982 [IORING_OP_OPENAT2] = {
984 [IORING_OP_EPOLL_CTL] = {
985 .unbound_nonreg_file = 1,
987 [IORING_OP_SPLICE] = {
990 .unbound_nonreg_file = 1,
992 [IORING_OP_PROVIDE_BUFFERS] = {},
993 [IORING_OP_REMOVE_BUFFERS] = {},
997 .unbound_nonreg_file = 1,
999 [IORING_OP_SHUTDOWN] = {
1002 [IORING_OP_RENAMEAT] = {},
1003 [IORING_OP_UNLINKAT] = {},
1006 static bool io_disarm_next(struct io_kiocb *req);
1007 static void io_uring_del_task_file(unsigned long index);
1008 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
1009 struct task_struct *task,
1010 struct files_struct *files);
1011 static void io_uring_cancel_sqpoll(struct io_ring_ctx *ctx);
1012 static void destroy_fixed_rsrc_ref_node(struct fixed_rsrc_ref_node *ref_node);
1013 static struct fixed_rsrc_ref_node *alloc_fixed_rsrc_ref_node(
1014 struct io_ring_ctx *ctx);
1015 static void io_ring_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc);
1017 static bool io_rw_reissue(struct io_kiocb *req);
1018 static void io_cqring_fill_event(struct io_kiocb *req, long res);
1019 static void io_put_req(struct io_kiocb *req);
1020 static void io_put_req_deferred(struct io_kiocb *req, int nr);
1021 static void io_double_put_req(struct io_kiocb *req);
1022 static void io_dismantle_req(struct io_kiocb *req);
1023 static void io_put_task(struct task_struct *task, int nr);
1024 static void io_queue_next(struct io_kiocb *req);
1025 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
1026 static void __io_queue_linked_timeout(struct io_kiocb *req);
1027 static void io_queue_linked_timeout(struct io_kiocb *req);
1028 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
1029 struct io_uring_rsrc_update *ip,
1031 static void __io_clean_op(struct io_kiocb *req);
1032 static struct file *io_file_get(struct io_submit_state *state,
1033 struct io_kiocb *req, int fd, bool fixed);
1034 static void __io_queue_sqe(struct io_kiocb *req);
1035 static void io_rsrc_put_work(struct work_struct *work);
1037 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
1038 struct iov_iter *iter, bool needs_lock);
1039 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
1040 const struct iovec *fast_iov,
1041 struct iov_iter *iter, bool force);
1042 static void io_req_task_queue(struct io_kiocb *req);
1043 static void io_submit_flush_completions(struct io_comp_state *cs,
1044 struct io_ring_ctx *ctx);
1046 static struct kmem_cache *req_cachep;
1048 static const struct file_operations io_uring_fops;
1050 struct sock *io_uring_get_socket(struct file *file)
1052 #if defined(CONFIG_UNIX)
1053 if (file->f_op == &io_uring_fops) {
1054 struct io_ring_ctx *ctx = file->private_data;
1056 return ctx->ring_sock->sk;
1061 EXPORT_SYMBOL(io_uring_get_socket);
1063 #define io_for_each_link(pos, head) \
1064 for (pos = (head); pos; pos = pos->link)
1066 static inline void io_clean_op(struct io_kiocb *req)
1068 if (req->flags & (REQ_F_NEED_CLEANUP | REQ_F_BUFFER_SELECTED))
1072 static inline void io_set_resource_node(struct io_kiocb *req)
1074 struct io_ring_ctx *ctx = req->ctx;
1076 if (!req->fixed_rsrc_refs) {
1077 req->fixed_rsrc_refs = &ctx->file_data->node->refs;
1078 percpu_ref_get(req->fixed_rsrc_refs);
1082 static bool io_match_task(struct io_kiocb *head,
1083 struct task_struct *task,
1084 struct files_struct *files)
1086 struct io_kiocb *req;
1088 if (task && head->task != task) {
1089 /* in terms of cancelation, always match if req task is dead */
1090 if (head->task->flags & PF_EXITING)
1097 io_for_each_link(req, head) {
1098 if (req->flags & REQ_F_INFLIGHT)
1104 static inline void req_set_fail_links(struct io_kiocb *req)
1106 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1107 req->flags |= REQ_F_FAIL_LINK;
1110 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1112 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1114 complete(&ctx->ref_comp);
1117 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1119 return !req->timeout.off;
1122 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1124 struct io_ring_ctx *ctx;
1127 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1132 * Use 5 bits less than the max cq entries, that should give us around
1133 * 32 entries per hash list if totally full and uniformly spread.
1135 hash_bits = ilog2(p->cq_entries);
1139 ctx->cancel_hash_bits = hash_bits;
1140 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1142 if (!ctx->cancel_hash)
1144 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1146 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1147 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1150 ctx->flags = p->flags;
1151 init_waitqueue_head(&ctx->sqo_sq_wait);
1152 INIT_LIST_HEAD(&ctx->sqd_list);
1153 init_waitqueue_head(&ctx->cq_wait);
1154 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1155 init_completion(&ctx->ref_comp);
1156 xa_init_flags(&ctx->io_buffers, XA_FLAGS_ALLOC1);
1157 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
1158 mutex_init(&ctx->uring_lock);
1159 init_waitqueue_head(&ctx->wait);
1160 spin_lock_init(&ctx->completion_lock);
1161 INIT_LIST_HEAD(&ctx->iopoll_list);
1162 INIT_LIST_HEAD(&ctx->defer_list);
1163 INIT_LIST_HEAD(&ctx->timeout_list);
1164 spin_lock_init(&ctx->inflight_lock);
1165 INIT_LIST_HEAD(&ctx->inflight_list);
1166 spin_lock_init(&ctx->rsrc_ref_lock);
1167 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
1168 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
1169 init_llist_head(&ctx->rsrc_put_llist);
1170 INIT_LIST_HEAD(&ctx->tctx_list);
1171 INIT_LIST_HEAD(&ctx->submit_state.comp.free_list);
1172 INIT_LIST_HEAD(&ctx->submit_state.comp.locked_free_list);
1175 kfree(ctx->cancel_hash);
1180 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1182 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1183 struct io_ring_ctx *ctx = req->ctx;
1185 return seq != ctx->cached_cq_tail
1186 + READ_ONCE(ctx->cached_cq_overflow);
1192 static void io_req_track_inflight(struct io_kiocb *req)
1194 struct io_ring_ctx *ctx = req->ctx;
1196 if (!(req->flags & REQ_F_INFLIGHT)) {
1197 req->flags |= REQ_F_INFLIGHT;
1199 spin_lock_irq(&ctx->inflight_lock);
1200 list_add(&req->inflight_entry, &ctx->inflight_list);
1201 spin_unlock_irq(&ctx->inflight_lock);
1205 static void io_prep_async_work(struct io_kiocb *req)
1207 const struct io_op_def *def = &io_op_defs[req->opcode];
1208 struct io_ring_ctx *ctx = req->ctx;
1210 if (!req->work.creds)
1211 req->work.creds = get_current_cred();
1213 if (req->flags & REQ_F_FORCE_ASYNC)
1214 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1216 if (req->flags & REQ_F_ISREG) {
1217 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1218 io_wq_hash_work(&req->work, file_inode(req->file));
1219 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
1220 if (def->unbound_nonreg_file)
1221 req->work.flags |= IO_WQ_WORK_UNBOUND;
1225 static void io_prep_async_link(struct io_kiocb *req)
1227 struct io_kiocb *cur;
1229 io_for_each_link(cur, req)
1230 io_prep_async_work(cur);
1233 static void io_queue_async_work(struct io_kiocb *req)
1235 struct io_ring_ctx *ctx = req->ctx;
1236 struct io_kiocb *link = io_prep_linked_timeout(req);
1237 struct io_uring_task *tctx = req->task->io_uring;
1240 BUG_ON(!tctx->io_wq);
1242 /* init ->work of the whole link before punting */
1243 io_prep_async_link(req);
1244 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1245 &req->work, req->flags);
1246 io_wq_enqueue(tctx->io_wq, &req->work);
1248 io_queue_linked_timeout(link);
1251 static void io_kill_timeout(struct io_kiocb *req, int status)
1253 struct io_timeout_data *io = req->async_data;
1256 ret = hrtimer_try_to_cancel(&io->timer);
1258 atomic_set(&req->ctx->cq_timeouts,
1259 atomic_read(&req->ctx->cq_timeouts) + 1);
1260 list_del_init(&req->timeout.list);
1261 io_cqring_fill_event(req, status);
1262 io_put_req_deferred(req, 1);
1266 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1269 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1270 struct io_defer_entry, list);
1272 if (req_need_defer(de->req, de->seq))
1274 list_del_init(&de->list);
1275 io_req_task_queue(de->req);
1277 } while (!list_empty(&ctx->defer_list));
1280 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1284 if (list_empty(&ctx->timeout_list))
1287 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1290 u32 events_needed, events_got;
1291 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1292 struct io_kiocb, timeout.list);
1294 if (io_is_timeout_noseq(req))
1298 * Since seq can easily wrap around over time, subtract
1299 * the last seq at which timeouts were flushed before comparing.
1300 * Assuming not more than 2^31-1 events have happened since,
1301 * these subtractions won't have wrapped, so we can check if
1302 * target is in [last_seq, current_seq] by comparing the two.
1304 events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1305 events_got = seq - ctx->cq_last_tm_flush;
1306 if (events_got < events_needed)
1309 list_del_init(&req->timeout.list);
1310 io_kill_timeout(req, 0);
1311 } while (!list_empty(&ctx->timeout_list));
1313 ctx->cq_last_tm_flush = seq;
1316 static void io_commit_cqring(struct io_ring_ctx *ctx)
1318 io_flush_timeouts(ctx);
1320 /* order cqe stores with ring update */
1321 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1323 if (unlikely(!list_empty(&ctx->defer_list)))
1324 __io_queue_deferred(ctx);
1327 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1329 struct io_rings *r = ctx->rings;
1331 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == r->sq_ring_entries;
1334 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1336 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1339 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1341 struct io_rings *rings = ctx->rings;
1345 * writes to the cq entry need to come after reading head; the
1346 * control dependency is enough as we're using WRITE_ONCE to
1349 if (__io_cqring_events(ctx) == rings->cq_ring_entries)
1352 tail = ctx->cached_cq_tail++;
1353 return &rings->cqes[tail & ctx->cq_mask];
1356 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1360 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1362 if (!ctx->eventfd_async)
1364 return io_wq_current_is_worker();
1367 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1369 /* see waitqueue_active() comment */
1372 if (waitqueue_active(&ctx->wait))
1373 wake_up(&ctx->wait);
1374 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1375 wake_up(&ctx->sq_data->wait);
1376 if (io_should_trigger_evfd(ctx))
1377 eventfd_signal(ctx->cq_ev_fd, 1);
1378 if (waitqueue_active(&ctx->cq_wait)) {
1379 wake_up_interruptible(&ctx->cq_wait);
1380 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1384 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1386 /* see waitqueue_active() comment */
1389 if (ctx->flags & IORING_SETUP_SQPOLL) {
1390 if (waitqueue_active(&ctx->wait))
1391 wake_up(&ctx->wait);
1393 if (io_should_trigger_evfd(ctx))
1394 eventfd_signal(ctx->cq_ev_fd, 1);
1395 if (waitqueue_active(&ctx->cq_wait)) {
1396 wake_up_interruptible(&ctx->cq_wait);
1397 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1401 /* Returns true if there are no backlogged entries after the flush */
1402 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1403 struct task_struct *tsk,
1404 struct files_struct *files)
1406 struct io_rings *rings = ctx->rings;
1407 struct io_kiocb *req, *tmp;
1408 struct io_uring_cqe *cqe;
1409 unsigned long flags;
1410 bool all_flushed, posted;
1413 if (!force && __io_cqring_events(ctx) == rings->cq_ring_entries)
1417 spin_lock_irqsave(&ctx->completion_lock, flags);
1418 list_for_each_entry_safe(req, tmp, &ctx->cq_overflow_list, compl.list) {
1419 if (!io_match_task(req, tsk, files))
1422 cqe = io_get_cqring(ctx);
1426 list_move(&req->compl.list, &list);
1428 WRITE_ONCE(cqe->user_data, req->user_data);
1429 WRITE_ONCE(cqe->res, req->result);
1430 WRITE_ONCE(cqe->flags, req->compl.cflags);
1432 ctx->cached_cq_overflow++;
1433 WRITE_ONCE(ctx->rings->cq_overflow,
1434 ctx->cached_cq_overflow);
1439 all_flushed = list_empty(&ctx->cq_overflow_list);
1441 clear_bit(0, &ctx->sq_check_overflow);
1442 clear_bit(0, &ctx->cq_check_overflow);
1443 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1447 io_commit_cqring(ctx);
1448 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1450 io_cqring_ev_posted(ctx);
1452 while (!list_empty(&list)) {
1453 req = list_first_entry(&list, struct io_kiocb, compl.list);
1454 list_del(&req->compl.list);
1461 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1462 struct task_struct *tsk,
1463 struct files_struct *files)
1467 if (test_bit(0, &ctx->cq_check_overflow)) {
1468 /* iopoll syncs against uring_lock, not completion_lock */
1469 if (ctx->flags & IORING_SETUP_IOPOLL)
1470 mutex_lock(&ctx->uring_lock);
1471 ret = __io_cqring_overflow_flush(ctx, force, tsk, files);
1472 if (ctx->flags & IORING_SETUP_IOPOLL)
1473 mutex_unlock(&ctx->uring_lock);
1479 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1481 struct io_ring_ctx *ctx = req->ctx;
1482 struct io_uring_cqe *cqe;
1484 trace_io_uring_complete(ctx, req->user_data, res);
1487 * If we can't get a cq entry, userspace overflowed the
1488 * submission (by quite a lot). Increment the overflow count in
1491 cqe = io_get_cqring(ctx);
1493 WRITE_ONCE(cqe->user_data, req->user_data);
1494 WRITE_ONCE(cqe->res, res);
1495 WRITE_ONCE(cqe->flags, cflags);
1496 } else if (ctx->cq_overflow_flushed ||
1497 atomic_read(&req->task->io_uring->in_idle)) {
1499 * If we're in ring overflow flush mode, or in task cancel mode,
1500 * then we cannot store the request for later flushing, we need
1501 * to drop it on the floor.
1503 ctx->cached_cq_overflow++;
1504 WRITE_ONCE(ctx->rings->cq_overflow, ctx->cached_cq_overflow);
1506 if (list_empty(&ctx->cq_overflow_list)) {
1507 set_bit(0, &ctx->sq_check_overflow);
1508 set_bit(0, &ctx->cq_check_overflow);
1509 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1513 req->compl.cflags = cflags;
1514 refcount_inc(&req->refs);
1515 list_add_tail(&req->compl.list, &ctx->cq_overflow_list);
1519 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1521 __io_cqring_fill_event(req, res, 0);
1524 static void io_req_complete_post(struct io_kiocb *req, long res,
1525 unsigned int cflags)
1527 struct io_ring_ctx *ctx = req->ctx;
1528 unsigned long flags;
1530 spin_lock_irqsave(&ctx->completion_lock, flags);
1531 __io_cqring_fill_event(req, res, cflags);
1533 * If we're the last reference to this request, add to our locked
1536 if (refcount_dec_and_test(&req->refs)) {
1537 struct io_comp_state *cs = &ctx->submit_state.comp;
1539 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
1540 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL_LINK))
1541 io_disarm_next(req);
1543 io_req_task_queue(req->link);
1547 io_dismantle_req(req);
1548 io_put_task(req->task, 1);
1549 list_add(&req->compl.list, &cs->locked_free_list);
1550 cs->locked_free_nr++;
1552 if (!percpu_ref_tryget(&ctx->refs))
1555 io_commit_cqring(ctx);
1556 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1559 io_cqring_ev_posted(ctx);
1560 percpu_ref_put(&ctx->refs);
1564 static void io_req_complete_state(struct io_kiocb *req, long res,
1565 unsigned int cflags)
1569 req->compl.cflags = cflags;
1570 req->flags |= REQ_F_COMPLETE_INLINE;
1573 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1574 long res, unsigned cflags)
1576 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1577 io_req_complete_state(req, res, cflags);
1579 io_req_complete_post(req, res, cflags);
1582 static inline void io_req_complete(struct io_kiocb *req, long res)
1584 __io_req_complete(req, 0, res, 0);
1587 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1589 struct io_submit_state *state = &ctx->submit_state;
1590 struct io_comp_state *cs = &state->comp;
1591 struct io_kiocb *req = NULL;
1594 * If we have more than a batch's worth of requests in our IRQ side
1595 * locked cache, grab the lock and move them over to our submission
1598 if (READ_ONCE(cs->locked_free_nr) > IO_COMPL_BATCH) {
1599 spin_lock_irq(&ctx->completion_lock);
1600 list_splice_init(&cs->locked_free_list, &cs->free_list);
1601 cs->locked_free_nr = 0;
1602 spin_unlock_irq(&ctx->completion_lock);
1605 while (!list_empty(&cs->free_list)) {
1606 req = list_first_entry(&cs->free_list, struct io_kiocb,
1608 list_del(&req->compl.list);
1609 state->reqs[state->free_reqs++] = req;
1610 if (state->free_reqs == ARRAY_SIZE(state->reqs))
1617 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1619 struct io_submit_state *state = &ctx->submit_state;
1621 BUILD_BUG_ON(IO_REQ_ALLOC_BATCH > ARRAY_SIZE(state->reqs));
1623 if (!state->free_reqs) {
1624 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1627 if (io_flush_cached_reqs(ctx))
1630 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1634 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1635 * retry single alloc to be on the safe side.
1637 if (unlikely(ret <= 0)) {
1638 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1639 if (!state->reqs[0])
1643 state->free_reqs = ret;
1647 return state->reqs[state->free_reqs];
1650 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1657 static void io_dismantle_req(struct io_kiocb *req)
1661 if (req->async_data)
1662 kfree(req->async_data);
1664 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1665 if (req->fixed_rsrc_refs)
1666 percpu_ref_put(req->fixed_rsrc_refs);
1667 if (req->work.creds) {
1668 put_cred(req->work.creds);
1669 req->work.creds = NULL;
1672 if (req->flags & REQ_F_INFLIGHT) {
1673 struct io_ring_ctx *ctx = req->ctx;
1674 unsigned long flags;
1676 spin_lock_irqsave(&ctx->inflight_lock, flags);
1677 list_del(&req->inflight_entry);
1678 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1679 req->flags &= ~REQ_F_INFLIGHT;
1683 /* must to be called somewhat shortly after putting a request */
1684 static inline void io_put_task(struct task_struct *task, int nr)
1686 struct io_uring_task *tctx = task->io_uring;
1688 percpu_counter_sub(&tctx->inflight, nr);
1689 if (unlikely(atomic_read(&tctx->in_idle)))
1690 wake_up(&tctx->wait);
1691 put_task_struct_many(task, nr);
1694 static void __io_free_req(struct io_kiocb *req)
1696 struct io_ring_ctx *ctx = req->ctx;
1698 io_dismantle_req(req);
1699 io_put_task(req->task, 1);
1701 kmem_cache_free(req_cachep, req);
1702 percpu_ref_put(&ctx->refs);
1705 static inline void io_remove_next_linked(struct io_kiocb *req)
1707 struct io_kiocb *nxt = req->link;
1709 req->link = nxt->link;
1713 static bool io_kill_linked_timeout(struct io_kiocb *req)
1714 __must_hold(&req->ctx->completion_lock)
1716 struct io_kiocb *link = req->link;
1717 bool cancelled = false;
1720 * Can happen if a linked timeout fired and link had been like
1721 * req -> link t-out -> link t-out [-> ...]
1723 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1724 struct io_timeout_data *io = link->async_data;
1727 io_remove_next_linked(req);
1728 link->timeout.head = NULL;
1729 ret = hrtimer_try_to_cancel(&io->timer);
1731 io_cqring_fill_event(link, -ECANCELED);
1732 io_put_req_deferred(link, 1);
1736 req->flags &= ~REQ_F_LINK_TIMEOUT;
1740 static void io_fail_links(struct io_kiocb *req)
1741 __must_hold(&req->ctx->completion_lock)
1743 struct io_kiocb *nxt, *link = req->link;
1750 trace_io_uring_fail_link(req, link);
1751 io_cqring_fill_event(link, -ECANCELED);
1752 io_put_req_deferred(link, 2);
1757 static bool io_disarm_next(struct io_kiocb *req)
1758 __must_hold(&req->ctx->completion_lock)
1760 bool posted = false;
1762 if (likely(req->flags & REQ_F_LINK_TIMEOUT))
1763 posted = io_kill_linked_timeout(req);
1764 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
1765 posted |= (req->link != NULL);
1771 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1773 struct io_kiocb *nxt;
1776 * If LINK is set, we have dependent requests in this chain. If we
1777 * didn't fail this request, queue the first one up, moving any other
1778 * dependencies to the next request. In case of failure, fail the rest
1781 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL_LINK)) {
1782 struct io_ring_ctx *ctx = req->ctx;
1783 unsigned long flags;
1786 spin_lock_irqsave(&ctx->completion_lock, flags);
1787 posted = io_disarm_next(req);
1789 io_commit_cqring(req->ctx);
1790 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1792 io_cqring_ev_posted(ctx);
1799 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1801 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
1803 return __io_req_find_next(req);
1806 static void ctx_flush_and_put(struct io_ring_ctx *ctx)
1810 if (ctx->submit_state.comp.nr) {
1811 mutex_lock(&ctx->uring_lock);
1812 io_submit_flush_completions(&ctx->submit_state.comp, ctx);
1813 mutex_unlock(&ctx->uring_lock);
1815 percpu_ref_put(&ctx->refs);
1818 static bool __tctx_task_work(struct io_uring_task *tctx)
1820 struct io_ring_ctx *ctx = NULL;
1821 struct io_wq_work_list list;
1822 struct io_wq_work_node *node;
1824 if (wq_list_empty(&tctx->task_list))
1827 spin_lock_irq(&tctx->task_lock);
1828 list = tctx->task_list;
1829 INIT_WQ_LIST(&tctx->task_list);
1830 spin_unlock_irq(&tctx->task_lock);
1834 struct io_wq_work_node *next = node->next;
1835 struct io_kiocb *req;
1837 req = container_of(node, struct io_kiocb, io_task_work.node);
1838 if (req->ctx != ctx) {
1839 ctx_flush_and_put(ctx);
1841 percpu_ref_get(&ctx->refs);
1844 req->task_work.func(&req->task_work);
1848 ctx_flush_and_put(ctx);
1849 return list.first != NULL;
1852 static void tctx_task_work(struct callback_head *cb)
1854 struct io_uring_task *tctx = container_of(cb, struct io_uring_task, task_work);
1856 clear_bit(0, &tctx->task_state);
1858 while (__tctx_task_work(tctx))
1862 static int io_task_work_add(struct task_struct *tsk, struct io_kiocb *req,
1863 enum task_work_notify_mode notify)
1865 struct io_uring_task *tctx = tsk->io_uring;
1866 struct io_wq_work_node *node, *prev;
1867 unsigned long flags;
1870 WARN_ON_ONCE(!tctx);
1872 spin_lock_irqsave(&tctx->task_lock, flags);
1873 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
1874 spin_unlock_irqrestore(&tctx->task_lock, flags);
1876 /* task_work already pending, we're done */
1877 if (test_bit(0, &tctx->task_state) ||
1878 test_and_set_bit(0, &tctx->task_state))
1881 if (!task_work_add(tsk, &tctx->task_work, notify))
1885 * Slow path - we failed, find and delete work. if the work is not
1886 * in the list, it got run and we're fine.
1889 spin_lock_irqsave(&tctx->task_lock, flags);
1890 wq_list_for_each(node, prev, &tctx->task_list) {
1891 if (&req->io_task_work.node == node) {
1892 wq_list_del(&tctx->task_list, node, prev);
1897 spin_unlock_irqrestore(&tctx->task_lock, flags);
1898 clear_bit(0, &tctx->task_state);
1902 static int io_req_task_work_add(struct io_kiocb *req)
1904 struct task_struct *tsk = req->task;
1905 struct io_ring_ctx *ctx = req->ctx;
1906 enum task_work_notify_mode notify;
1909 if (tsk->flags & PF_EXITING)
1913 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1914 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1915 * processing task_work. There's no reliable way to tell if TWA_RESUME
1919 if (!(ctx->flags & IORING_SETUP_SQPOLL))
1920 notify = TWA_SIGNAL;
1922 ret = io_task_work_add(tsk, req, notify);
1924 wake_up_process(tsk);
1929 static bool io_run_task_work_head(struct callback_head **work_head)
1931 struct callback_head *work, *next;
1932 bool executed = false;
1935 work = xchg(work_head, NULL);
1951 static void io_task_work_add_head(struct callback_head **work_head,
1952 struct callback_head *task_work)
1954 struct callback_head *head;
1957 head = READ_ONCE(*work_head);
1958 task_work->next = head;
1959 } while (cmpxchg(work_head, head, task_work) != head);
1962 static void io_req_task_work_add_fallback(struct io_kiocb *req,
1963 task_work_func_t cb)
1965 init_task_work(&req->task_work, cb);
1966 io_task_work_add_head(&req->ctx->exit_task_work, &req->task_work);
1969 static void __io_req_task_cancel(struct io_kiocb *req, int error)
1971 struct io_ring_ctx *ctx = req->ctx;
1973 spin_lock_irq(&ctx->completion_lock);
1974 io_cqring_fill_event(req, error);
1975 io_commit_cqring(ctx);
1976 spin_unlock_irq(&ctx->completion_lock);
1978 io_cqring_ev_posted(ctx);
1979 req_set_fail_links(req);
1980 io_double_put_req(req);
1983 static void io_req_task_cancel(struct callback_head *cb)
1985 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1986 struct io_ring_ctx *ctx = req->ctx;
1988 mutex_lock(&ctx->uring_lock);
1989 __io_req_task_cancel(req, req->result);
1990 mutex_unlock(&ctx->uring_lock);
1991 percpu_ref_put(&ctx->refs);
1994 static void __io_req_task_submit(struct io_kiocb *req)
1996 struct io_ring_ctx *ctx = req->ctx;
1998 /* ctx stays valid until unlock, even if we drop all ours ctx->refs */
1999 mutex_lock(&ctx->uring_lock);
2000 if (!(current->flags & PF_EXITING) && !current->in_execve)
2001 __io_queue_sqe(req);
2003 __io_req_task_cancel(req, -EFAULT);
2004 mutex_unlock(&ctx->uring_lock);
2007 static void io_req_task_submit(struct callback_head *cb)
2009 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2011 __io_req_task_submit(req);
2014 static void io_req_task_queue(struct io_kiocb *req)
2018 req->task_work.func = io_req_task_submit;
2019 ret = io_req_task_work_add(req);
2020 if (unlikely(ret)) {
2021 req->result = -ECANCELED;
2022 percpu_ref_get(&req->ctx->refs);
2023 io_req_task_work_add_fallback(req, io_req_task_cancel);
2027 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2029 percpu_ref_get(&req->ctx->refs);
2031 req->task_work.func = io_req_task_cancel;
2033 if (unlikely(io_req_task_work_add(req)))
2034 io_req_task_work_add_fallback(req, io_req_task_cancel);
2037 static inline void io_queue_next(struct io_kiocb *req)
2039 struct io_kiocb *nxt = io_req_find_next(req);
2042 io_req_task_queue(nxt);
2045 static void io_free_req(struct io_kiocb *req)
2052 struct task_struct *task;
2057 static inline void io_init_req_batch(struct req_batch *rb)
2064 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2065 struct req_batch *rb)
2068 io_put_task(rb->task, rb->task_refs);
2070 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2073 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2074 struct io_submit_state *state)
2078 if (req->task != rb->task) {
2080 io_put_task(rb->task, rb->task_refs);
2081 rb->task = req->task;
2087 io_dismantle_req(req);
2088 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2089 state->reqs[state->free_reqs++] = req;
2091 list_add(&req->compl.list, &state->comp.free_list);
2094 static void io_submit_flush_completions(struct io_comp_state *cs,
2095 struct io_ring_ctx *ctx)
2098 struct io_kiocb *req;
2099 struct req_batch rb;
2101 io_init_req_batch(&rb);
2102 spin_lock_irq(&ctx->completion_lock);
2103 for (i = 0; i < nr; i++) {
2105 __io_cqring_fill_event(req, req->result, req->compl.cflags);
2107 io_commit_cqring(ctx);
2108 spin_unlock_irq(&ctx->completion_lock);
2110 io_cqring_ev_posted(ctx);
2111 for (i = 0; i < nr; i++) {
2114 /* submission and completion refs */
2115 if (refcount_sub_and_test(2, &req->refs))
2116 io_req_free_batch(&rb, req, &ctx->submit_state);
2119 io_req_free_batch_finish(ctx, &rb);
2124 * Drop reference to request, return next in chain (if there is one) if this
2125 * was the last reference to this request.
2127 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2129 struct io_kiocb *nxt = NULL;
2131 if (refcount_dec_and_test(&req->refs)) {
2132 nxt = io_req_find_next(req);
2138 static void io_put_req(struct io_kiocb *req)
2140 if (refcount_dec_and_test(&req->refs))
2144 static void io_put_req_deferred_cb(struct callback_head *cb)
2146 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2151 static void io_free_req_deferred(struct io_kiocb *req)
2155 req->task_work.func = io_put_req_deferred_cb;
2156 ret = io_req_task_work_add(req);
2158 io_req_task_work_add_fallback(req, io_put_req_deferred_cb);
2161 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2163 if (refcount_sub_and_test(refs, &req->refs))
2164 io_free_req_deferred(req);
2167 static void io_double_put_req(struct io_kiocb *req)
2169 /* drop both submit and complete references */
2170 if (refcount_sub_and_test(2, &req->refs))
2174 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2176 /* See comment at the top of this file */
2178 return __io_cqring_events(ctx);
2181 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2183 struct io_rings *rings = ctx->rings;
2185 /* make sure SQ entry isn't read before tail */
2186 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2189 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2191 unsigned int cflags;
2193 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2194 cflags |= IORING_CQE_F_BUFFER;
2195 req->flags &= ~REQ_F_BUFFER_SELECTED;
2200 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2202 struct io_buffer *kbuf;
2204 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2205 return io_put_kbuf(req, kbuf);
2208 static inline bool io_run_task_work(void)
2211 * Not safe to run on exiting task, and the task_work handling will
2212 * not add work to such a task.
2214 if (unlikely(current->flags & PF_EXITING))
2216 if (current->task_works) {
2217 __set_current_state(TASK_RUNNING);
2226 * Find and free completed poll iocbs
2228 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2229 struct list_head *done)
2231 struct req_batch rb;
2232 struct io_kiocb *req;
2234 /* order with ->result store in io_complete_rw_iopoll() */
2237 io_init_req_batch(&rb);
2238 while (!list_empty(done)) {
2241 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2242 list_del(&req->inflight_entry);
2244 if (READ_ONCE(req->result) == -EAGAIN) {
2245 req->iopoll_completed = 0;
2246 if (io_rw_reissue(req))
2250 if (req->flags & REQ_F_BUFFER_SELECTED)
2251 cflags = io_put_rw_kbuf(req);
2253 __io_cqring_fill_event(req, req->result, cflags);
2256 if (refcount_dec_and_test(&req->refs))
2257 io_req_free_batch(&rb, req, &ctx->submit_state);
2260 io_commit_cqring(ctx);
2261 io_cqring_ev_posted_iopoll(ctx);
2262 io_req_free_batch_finish(ctx, &rb);
2265 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2268 struct io_kiocb *req, *tmp;
2274 * Only spin for completions if we don't have multiple devices hanging
2275 * off our complete list, and we're under the requested amount.
2277 spin = !ctx->poll_multi_file && *nr_events < min;
2280 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2281 struct kiocb *kiocb = &req->rw.kiocb;
2284 * Move completed and retryable entries to our local lists.
2285 * If we find a request that requires polling, break out
2286 * and complete those lists first, if we have entries there.
2288 if (READ_ONCE(req->iopoll_completed)) {
2289 list_move_tail(&req->inflight_entry, &done);
2292 if (!list_empty(&done))
2295 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2299 /* iopoll may have completed current req */
2300 if (READ_ONCE(req->iopoll_completed))
2301 list_move_tail(&req->inflight_entry, &done);
2308 if (!list_empty(&done))
2309 io_iopoll_complete(ctx, nr_events, &done);
2315 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2316 * non-spinning poll check - we'll still enter the driver poll loop, but only
2317 * as a non-spinning completion check.
2319 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2322 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2325 ret = io_do_iopoll(ctx, nr_events, min);
2328 if (*nr_events >= min)
2336 * We can't just wait for polled events to come to us, we have to actively
2337 * find and complete them.
2339 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2341 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2344 mutex_lock(&ctx->uring_lock);
2345 while (!list_empty(&ctx->iopoll_list)) {
2346 unsigned int nr_events = 0;
2348 io_do_iopoll(ctx, &nr_events, 0);
2350 /* let it sleep and repeat later if can't complete a request */
2354 * Ensure we allow local-to-the-cpu processing to take place,
2355 * in this case we need to ensure that we reap all events.
2356 * Also let task_work, etc. to progress by releasing the mutex
2358 if (need_resched()) {
2359 mutex_unlock(&ctx->uring_lock);
2361 mutex_lock(&ctx->uring_lock);
2364 mutex_unlock(&ctx->uring_lock);
2367 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2369 unsigned int nr_events = 0;
2370 int iters = 0, ret = 0;
2373 * We disallow the app entering submit/complete with polling, but we
2374 * still need to lock the ring to prevent racing with polled issue
2375 * that got punted to a workqueue.
2377 mutex_lock(&ctx->uring_lock);
2380 * Don't enter poll loop if we already have events pending.
2381 * If we do, we can potentially be spinning for commands that
2382 * already triggered a CQE (eg in error).
2384 if (test_bit(0, &ctx->cq_check_overflow))
2385 __io_cqring_overflow_flush(ctx, false, NULL, NULL);
2386 if (io_cqring_events(ctx))
2390 * If a submit got punted to a workqueue, we can have the
2391 * application entering polling for a command before it gets
2392 * issued. That app will hold the uring_lock for the duration
2393 * of the poll right here, so we need to take a breather every
2394 * now and then to ensure that the issue has a chance to add
2395 * the poll to the issued list. Otherwise we can spin here
2396 * forever, while the workqueue is stuck trying to acquire the
2399 if (!(++iters & 7)) {
2400 mutex_unlock(&ctx->uring_lock);
2402 mutex_lock(&ctx->uring_lock);
2405 ret = io_iopoll_getevents(ctx, &nr_events, min);
2409 } while (min && !nr_events && !need_resched());
2411 mutex_unlock(&ctx->uring_lock);
2415 static void kiocb_end_write(struct io_kiocb *req)
2418 * Tell lockdep we inherited freeze protection from submission
2421 if (req->flags & REQ_F_ISREG) {
2422 struct inode *inode = file_inode(req->file);
2424 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2426 file_end_write(req->file);
2430 static bool io_resubmit_prep(struct io_kiocb *req)
2432 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2434 struct iov_iter iter;
2436 /* already prepared */
2437 if (req->async_data)
2440 switch (req->opcode) {
2441 case IORING_OP_READV:
2442 case IORING_OP_READ_FIXED:
2443 case IORING_OP_READ:
2446 case IORING_OP_WRITEV:
2447 case IORING_OP_WRITE_FIXED:
2448 case IORING_OP_WRITE:
2452 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2457 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2460 return !io_setup_async_rw(req, iovec, inline_vecs, &iter, false);
2463 static bool io_rw_should_reissue(struct io_kiocb *req)
2465 umode_t mode = file_inode(req->file)->i_mode;
2466 struct io_ring_ctx *ctx = req->ctx;
2468 if (!S_ISBLK(mode) && !S_ISREG(mode))
2470 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2471 !(ctx->flags & IORING_SETUP_IOPOLL)))
2474 * If ref is dying, we might be running poll reap from the exit work.
2475 * Don't attempt to reissue from that path, just let it fail with
2478 if (percpu_ref_is_dying(&ctx->refs))
2484 static bool io_rw_reissue(struct io_kiocb *req)
2487 if (!io_rw_should_reissue(req))
2490 lockdep_assert_held(&req->ctx->uring_lock);
2492 if (io_resubmit_prep(req)) {
2493 refcount_inc(&req->refs);
2494 io_queue_async_work(req);
2497 req_set_fail_links(req);
2502 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2503 unsigned int issue_flags)
2507 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2508 kiocb_end_write(req);
2509 if ((res == -EAGAIN || res == -EOPNOTSUPP) && io_rw_should_reissue(req)) {
2510 req->flags |= REQ_F_REISSUE;
2513 if (res != req->result)
2514 req_set_fail_links(req);
2515 if (req->flags & REQ_F_BUFFER_SELECTED)
2516 cflags = io_put_rw_kbuf(req);
2517 __io_req_complete(req, issue_flags, res, cflags);
2520 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2522 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2524 __io_complete_rw(req, res, res2, 0);
2527 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2529 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2532 /* Rewind iter, if we have one. iopoll path resubmits as usual */
2533 if (res == -EAGAIN && io_rw_should_reissue(req)) {
2534 struct io_async_rw *rw = req->async_data;
2537 iov_iter_revert(&rw->iter,
2538 req->result - iov_iter_count(&rw->iter));
2539 else if (!io_resubmit_prep(req))
2544 if (kiocb->ki_flags & IOCB_WRITE)
2545 kiocb_end_write(req);
2547 if (res != -EAGAIN && res != req->result)
2548 req_set_fail_links(req);
2550 WRITE_ONCE(req->result, res);
2551 /* order with io_poll_complete() checking ->result */
2553 WRITE_ONCE(req->iopoll_completed, 1);
2557 * After the iocb has been issued, it's safe to be found on the poll list.
2558 * Adding the kiocb to the list AFTER submission ensures that we don't
2559 * find it from a io_iopoll_getevents() thread before the issuer is done
2560 * accessing the kiocb cookie.
2562 static void io_iopoll_req_issued(struct io_kiocb *req, bool in_async)
2564 struct io_ring_ctx *ctx = req->ctx;
2567 * Track whether we have multiple files in our lists. This will impact
2568 * how we do polling eventually, not spinning if we're on potentially
2569 * different devices.
2571 if (list_empty(&ctx->iopoll_list)) {
2572 ctx->poll_multi_file = false;
2573 } else if (!ctx->poll_multi_file) {
2574 struct io_kiocb *list_req;
2576 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2578 if (list_req->file != req->file)
2579 ctx->poll_multi_file = true;
2583 * For fast devices, IO may have already completed. If it has, add
2584 * it to the front so we find it first.
2586 if (READ_ONCE(req->iopoll_completed))
2587 list_add(&req->inflight_entry, &ctx->iopoll_list);
2589 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2592 * If IORING_SETUP_SQPOLL is enabled, sqes are either handled in sq thread
2593 * task context or in io worker task context. If current task context is
2594 * sq thread, we don't need to check whether should wake up sq thread.
2596 if (in_async && (ctx->flags & IORING_SETUP_SQPOLL) &&
2597 wq_has_sleeper(&ctx->sq_data->wait))
2598 wake_up(&ctx->sq_data->wait);
2601 static inline void io_state_file_put(struct io_submit_state *state)
2603 if (state->file_refs) {
2604 fput_many(state->file, state->file_refs);
2605 state->file_refs = 0;
2610 * Get as many references to a file as we have IOs left in this submission,
2611 * assuming most submissions are for one file, or at least that each file
2612 * has more than one submission.
2614 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2619 if (state->file_refs) {
2620 if (state->fd == fd) {
2624 io_state_file_put(state);
2626 state->file = fget_many(fd, state->ios_left);
2627 if (unlikely(!state->file))
2631 state->file_refs = state->ios_left - 1;
2635 static bool io_bdev_nowait(struct block_device *bdev)
2637 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2641 * If we tracked the file through the SCM inflight mechanism, we could support
2642 * any file. For now, just ensure that anything potentially problematic is done
2645 static bool io_file_supports_async(struct file *file, int rw)
2647 umode_t mode = file_inode(file)->i_mode;
2649 if (S_ISBLK(mode)) {
2650 if (IS_ENABLED(CONFIG_BLOCK) &&
2651 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2655 if (S_ISCHR(mode) || S_ISSOCK(mode))
2657 if (S_ISREG(mode)) {
2658 if (IS_ENABLED(CONFIG_BLOCK) &&
2659 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2660 file->f_op != &io_uring_fops)
2665 /* any ->read/write should understand O_NONBLOCK */
2666 if (file->f_flags & O_NONBLOCK)
2669 if (!(file->f_mode & FMODE_NOWAIT))
2673 return file->f_op->read_iter != NULL;
2675 return file->f_op->write_iter != NULL;
2678 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2680 struct io_ring_ctx *ctx = req->ctx;
2681 struct kiocb *kiocb = &req->rw.kiocb;
2682 struct file *file = req->file;
2686 if (S_ISREG(file_inode(file)->i_mode))
2687 req->flags |= REQ_F_ISREG;
2689 kiocb->ki_pos = READ_ONCE(sqe->off);
2690 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2691 req->flags |= REQ_F_CUR_POS;
2692 kiocb->ki_pos = file->f_pos;
2694 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2695 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2696 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2700 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2701 if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2702 req->flags |= REQ_F_NOWAIT;
2704 ioprio = READ_ONCE(sqe->ioprio);
2706 ret = ioprio_check_cap(ioprio);
2710 kiocb->ki_ioprio = ioprio;
2712 kiocb->ki_ioprio = get_current_ioprio();
2714 if (ctx->flags & IORING_SETUP_IOPOLL) {
2715 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2716 !kiocb->ki_filp->f_op->iopoll)
2719 kiocb->ki_flags |= IOCB_HIPRI;
2720 kiocb->ki_complete = io_complete_rw_iopoll;
2721 req->iopoll_completed = 0;
2723 if (kiocb->ki_flags & IOCB_HIPRI)
2725 kiocb->ki_complete = io_complete_rw;
2728 req->rw.addr = READ_ONCE(sqe->addr);
2729 req->rw.len = READ_ONCE(sqe->len);
2730 req->buf_index = READ_ONCE(sqe->buf_index);
2734 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2740 case -ERESTARTNOINTR:
2741 case -ERESTARTNOHAND:
2742 case -ERESTART_RESTARTBLOCK:
2744 * We can't just restart the syscall, since previously
2745 * submitted sqes may already be in progress. Just fail this
2751 kiocb->ki_complete(kiocb, ret, 0);
2755 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2756 unsigned int issue_flags)
2758 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2759 struct io_async_rw *io = req->async_data;
2761 /* add previously done IO, if any */
2762 if (io && io->bytes_done > 0) {
2764 ret = io->bytes_done;
2766 ret += io->bytes_done;
2769 if (req->flags & REQ_F_CUR_POS)
2770 req->file->f_pos = kiocb->ki_pos;
2771 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2772 __io_complete_rw(req, ret, 0, issue_flags);
2774 io_rw_done(kiocb, ret);
2777 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
2779 struct io_ring_ctx *ctx = req->ctx;
2780 size_t len = req->rw.len;
2781 struct io_mapped_ubuf *imu;
2782 u16 index, buf_index = req->buf_index;
2786 if (unlikely(buf_index >= ctx->nr_user_bufs))
2788 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2789 imu = &ctx->user_bufs[index];
2790 buf_addr = req->rw.addr;
2793 if (buf_addr + len < buf_addr)
2795 /* not inside the mapped region */
2796 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2800 * May not be a start of buffer, set size appropriately
2801 * and advance us to the beginning.
2803 offset = buf_addr - imu->ubuf;
2804 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2808 * Don't use iov_iter_advance() here, as it's really slow for
2809 * using the latter parts of a big fixed buffer - it iterates
2810 * over each segment manually. We can cheat a bit here, because
2813 * 1) it's a BVEC iter, we set it up
2814 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2815 * first and last bvec
2817 * So just find our index, and adjust the iterator afterwards.
2818 * If the offset is within the first bvec (or the whole first
2819 * bvec, just use iov_iter_advance(). This makes it easier
2820 * since we can just skip the first segment, which may not
2821 * be PAGE_SIZE aligned.
2823 const struct bio_vec *bvec = imu->bvec;
2825 if (offset <= bvec->bv_len) {
2826 iov_iter_advance(iter, offset);
2828 unsigned long seg_skip;
2830 /* skip first vec */
2831 offset -= bvec->bv_len;
2832 seg_skip = 1 + (offset >> PAGE_SHIFT);
2834 iter->bvec = bvec + seg_skip;
2835 iter->nr_segs -= seg_skip;
2836 iter->count -= bvec->bv_len + offset;
2837 iter->iov_offset = offset & ~PAGE_MASK;
2844 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2847 mutex_unlock(&ctx->uring_lock);
2850 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2853 * "Normal" inline submissions always hold the uring_lock, since we
2854 * grab it from the system call. Same is true for the SQPOLL offload.
2855 * The only exception is when we've detached the request and issue it
2856 * from an async worker thread, grab the lock for that case.
2859 mutex_lock(&ctx->uring_lock);
2862 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2863 int bgid, struct io_buffer *kbuf,
2866 struct io_buffer *head;
2868 if (req->flags & REQ_F_BUFFER_SELECTED)
2871 io_ring_submit_lock(req->ctx, needs_lock);
2873 lockdep_assert_held(&req->ctx->uring_lock);
2875 head = xa_load(&req->ctx->io_buffers, bgid);
2877 if (!list_empty(&head->list)) {
2878 kbuf = list_last_entry(&head->list, struct io_buffer,
2880 list_del(&kbuf->list);
2883 xa_erase(&req->ctx->io_buffers, bgid);
2885 if (*len > kbuf->len)
2888 kbuf = ERR_PTR(-ENOBUFS);
2891 io_ring_submit_unlock(req->ctx, needs_lock);
2896 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2899 struct io_buffer *kbuf;
2902 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2903 bgid = req->buf_index;
2904 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2907 req->rw.addr = (u64) (unsigned long) kbuf;
2908 req->flags |= REQ_F_BUFFER_SELECTED;
2909 return u64_to_user_ptr(kbuf->addr);
2912 #ifdef CONFIG_COMPAT
2913 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2916 struct compat_iovec __user *uiov;
2917 compat_ssize_t clen;
2921 uiov = u64_to_user_ptr(req->rw.addr);
2922 if (!access_ok(uiov, sizeof(*uiov)))
2924 if (__get_user(clen, &uiov->iov_len))
2930 buf = io_rw_buffer_select(req, &len, needs_lock);
2932 return PTR_ERR(buf);
2933 iov[0].iov_base = buf;
2934 iov[0].iov_len = (compat_size_t) len;
2939 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2942 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2946 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2949 len = iov[0].iov_len;
2952 buf = io_rw_buffer_select(req, &len, needs_lock);
2954 return PTR_ERR(buf);
2955 iov[0].iov_base = buf;
2956 iov[0].iov_len = len;
2960 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2963 if (req->flags & REQ_F_BUFFER_SELECTED) {
2964 struct io_buffer *kbuf;
2966 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2967 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2968 iov[0].iov_len = kbuf->len;
2971 if (req->rw.len != 1)
2974 #ifdef CONFIG_COMPAT
2975 if (req->ctx->compat)
2976 return io_compat_import(req, iov, needs_lock);
2979 return __io_iov_buffer_select(req, iov, needs_lock);
2982 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
2983 struct iov_iter *iter, bool needs_lock)
2985 void __user *buf = u64_to_user_ptr(req->rw.addr);
2986 size_t sqe_len = req->rw.len;
2987 u8 opcode = req->opcode;
2990 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2992 return io_import_fixed(req, rw, iter);
2995 /* buffer index only valid with fixed read/write, or buffer select */
2996 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
2999 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3000 if (req->flags & REQ_F_BUFFER_SELECT) {
3001 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3003 return PTR_ERR(buf);
3004 req->rw.len = sqe_len;
3007 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3012 if (req->flags & REQ_F_BUFFER_SELECT) {
3013 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3015 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3020 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3024 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3026 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3030 * For files that don't have ->read_iter() and ->write_iter(), handle them
3031 * by looping over ->read() or ->write() manually.
3033 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3035 struct kiocb *kiocb = &req->rw.kiocb;
3036 struct file *file = req->file;
3040 * Don't support polled IO through this interface, and we can't
3041 * support non-blocking either. For the latter, this just causes
3042 * the kiocb to be handled from an async context.
3044 if (kiocb->ki_flags & IOCB_HIPRI)
3046 if (kiocb->ki_flags & IOCB_NOWAIT)
3049 while (iov_iter_count(iter)) {
3053 if (!iov_iter_is_bvec(iter)) {
3054 iovec = iov_iter_iovec(iter);
3056 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3057 iovec.iov_len = req->rw.len;
3061 nr = file->f_op->read(file, iovec.iov_base,
3062 iovec.iov_len, io_kiocb_ppos(kiocb));
3064 nr = file->f_op->write(file, iovec.iov_base,
3065 iovec.iov_len, io_kiocb_ppos(kiocb));
3074 if (nr != iovec.iov_len)
3078 iov_iter_advance(iter, nr);
3084 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3085 const struct iovec *fast_iov, struct iov_iter *iter)
3087 struct io_async_rw *rw = req->async_data;
3089 memcpy(&rw->iter, iter, sizeof(*iter));
3090 rw->free_iovec = iovec;
3092 /* can only be fixed buffers, no need to do anything */
3093 if (iov_iter_is_bvec(iter))
3096 unsigned iov_off = 0;
3098 rw->iter.iov = rw->fast_iov;
3099 if (iter->iov != fast_iov) {
3100 iov_off = iter->iov - fast_iov;
3101 rw->iter.iov += iov_off;
3103 if (rw->fast_iov != fast_iov)
3104 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3105 sizeof(struct iovec) * iter->nr_segs);
3107 req->flags |= REQ_F_NEED_CLEANUP;
3111 static inline int __io_alloc_async_data(struct io_kiocb *req)
3113 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3114 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3115 return req->async_data == NULL;
3118 static int io_alloc_async_data(struct io_kiocb *req)
3120 if (!io_op_defs[req->opcode].needs_async_data)
3123 return __io_alloc_async_data(req);
3126 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3127 const struct iovec *fast_iov,
3128 struct iov_iter *iter, bool force)
3130 if (!force && !io_op_defs[req->opcode].needs_async_data)
3132 if (!req->async_data) {
3133 if (__io_alloc_async_data(req)) {
3138 io_req_map_rw(req, iovec, fast_iov, iter);
3143 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3145 struct io_async_rw *iorw = req->async_data;
3146 struct iovec *iov = iorw->fast_iov;
3149 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3150 if (unlikely(ret < 0))
3153 iorw->bytes_done = 0;
3154 iorw->free_iovec = iov;
3156 req->flags |= REQ_F_NEED_CLEANUP;
3160 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3162 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3164 return io_prep_rw(req, sqe);
3168 * This is our waitqueue callback handler, registered through lock_page_async()
3169 * when we initially tried to do the IO with the iocb armed our waitqueue.
3170 * This gets called when the page is unlocked, and we generally expect that to
3171 * happen when the page IO is completed and the page is now uptodate. This will
3172 * queue a task_work based retry of the operation, attempting to copy the data
3173 * again. If the latter fails because the page was NOT uptodate, then we will
3174 * do a thread based blocking retry of the operation. That's the unexpected
3177 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3178 int sync, void *arg)
3180 struct wait_page_queue *wpq;
3181 struct io_kiocb *req = wait->private;
3182 struct wait_page_key *key = arg;
3184 wpq = container_of(wait, struct wait_page_queue, wait);
3186 if (!wake_page_match(wpq, key))
3189 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3190 list_del_init(&wait->entry);
3192 /* submit ref gets dropped, acquire a new one */
3193 refcount_inc(&req->refs);
3194 io_req_task_queue(req);
3199 * This controls whether a given IO request should be armed for async page
3200 * based retry. If we return false here, the request is handed to the async
3201 * worker threads for retry. If we're doing buffered reads on a regular file,
3202 * we prepare a private wait_page_queue entry and retry the operation. This
3203 * will either succeed because the page is now uptodate and unlocked, or it
3204 * will register a callback when the page is unlocked at IO completion. Through
3205 * that callback, io_uring uses task_work to setup a retry of the operation.
3206 * That retry will attempt the buffered read again. The retry will generally
3207 * succeed, or in rare cases where it fails, we then fall back to using the
3208 * async worker threads for a blocking retry.
3210 static bool io_rw_should_retry(struct io_kiocb *req)
3212 struct io_async_rw *rw = req->async_data;
3213 struct wait_page_queue *wait = &rw->wpq;
3214 struct kiocb *kiocb = &req->rw.kiocb;
3216 /* never retry for NOWAIT, we just complete with -EAGAIN */
3217 if (req->flags & REQ_F_NOWAIT)
3220 /* Only for buffered IO */
3221 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3225 * just use poll if we can, and don't attempt if the fs doesn't
3226 * support callback based unlocks
3228 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3231 wait->wait.func = io_async_buf_func;
3232 wait->wait.private = req;
3233 wait->wait.flags = 0;
3234 INIT_LIST_HEAD(&wait->wait.entry);
3235 kiocb->ki_flags |= IOCB_WAITQ;
3236 kiocb->ki_flags &= ~IOCB_NOWAIT;
3237 kiocb->ki_waitq = wait;
3241 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3243 if (req->file->f_op->read_iter)
3244 return call_read_iter(req->file, &req->rw.kiocb, iter);
3245 else if (req->file->f_op->read)
3246 return loop_rw_iter(READ, req, iter);
3251 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3253 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3254 struct kiocb *kiocb = &req->rw.kiocb;
3255 struct iov_iter __iter, *iter = &__iter;
3256 struct io_async_rw *rw = req->async_data;
3257 ssize_t io_size, ret, ret2;
3258 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3264 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3268 io_size = iov_iter_count(iter);
3269 req->result = io_size;
3271 /* Ensure we clear previously set non-block flag */
3272 if (!force_nonblock)
3273 kiocb->ki_flags &= ~IOCB_NOWAIT;
3275 kiocb->ki_flags |= IOCB_NOWAIT;
3277 /* If the file doesn't support async, just async punt */
3278 if (force_nonblock && !io_file_supports_async(req->file, READ)) {
3279 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3280 return ret ?: -EAGAIN;
3283 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3284 if (unlikely(ret)) {
3289 ret = io_iter_do_read(req, iter);
3291 if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
3292 /* IOPOLL retry should happen for io-wq threads */
3293 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3295 /* no retry on NONBLOCK nor RWF_NOWAIT */
3296 if (req->flags & REQ_F_NOWAIT)
3298 /* some cases will consume bytes even on error returns */
3299 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3301 } else if (ret == -EIOCBQUEUED) {
3303 } else if (ret <= 0 || ret == io_size || !force_nonblock ||
3304 (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3305 /* read all, failed, already did sync or don't want to retry */
3309 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3314 rw = req->async_data;
3315 /* now use our persistent iterator, if we aren't already */
3320 rw->bytes_done += ret;
3321 /* if we can retry, do so with the callbacks armed */
3322 if (!io_rw_should_retry(req)) {
3323 kiocb->ki_flags &= ~IOCB_WAITQ;
3328 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3329 * we get -EIOCBQUEUED, then we'll get a notification when the
3330 * desired page gets unlocked. We can also get a partial read
3331 * here, and if we do, then just retry at the new offset.
3333 ret = io_iter_do_read(req, iter);
3334 if (ret == -EIOCBQUEUED)
3336 /* we got some bytes, but not all. retry. */
3337 kiocb->ki_flags &= ~IOCB_WAITQ;
3338 } while (ret > 0 && ret < io_size);
3340 kiocb_done(kiocb, ret, issue_flags);
3342 /* it's faster to check here then delegate to kfree */
3348 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3350 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3352 return io_prep_rw(req, sqe);
3355 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3357 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3358 struct kiocb *kiocb = &req->rw.kiocb;
3359 struct iov_iter __iter, *iter = &__iter;
3360 struct io_async_rw *rw = req->async_data;
3361 ssize_t ret, ret2, io_size;
3362 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3368 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3372 io_size = iov_iter_count(iter);
3373 req->result = io_size;
3375 /* Ensure we clear previously set non-block flag */
3376 if (!force_nonblock)
3377 kiocb->ki_flags &= ~IOCB_NOWAIT;
3379 kiocb->ki_flags |= IOCB_NOWAIT;
3381 /* If the file doesn't support async, just async punt */
3382 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3385 /* file path doesn't support NOWAIT for non-direct_IO */
3386 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3387 (req->flags & REQ_F_ISREG))
3390 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3395 * Open-code file_start_write here to grab freeze protection,
3396 * which will be released by another thread in
3397 * io_complete_rw(). Fool lockdep by telling it the lock got
3398 * released so that it doesn't complain about the held lock when
3399 * we return to userspace.
3401 if (req->flags & REQ_F_ISREG) {
3402 sb_start_write(file_inode(req->file)->i_sb);
3403 __sb_writers_release(file_inode(req->file)->i_sb,
3406 kiocb->ki_flags |= IOCB_WRITE;
3408 if (req->file->f_op->write_iter)
3409 ret2 = call_write_iter(req->file, kiocb, iter);
3410 else if (req->file->f_op->write)
3411 ret2 = loop_rw_iter(WRITE, req, iter);
3415 if (req->flags & REQ_F_REISSUE)
3419 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3420 * retry them without IOCB_NOWAIT.
3422 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3424 /* no retry on NONBLOCK nor RWF_NOWAIT */
3425 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3427 if (!force_nonblock || ret2 != -EAGAIN) {
3428 /* IOPOLL retry should happen for io-wq threads */
3429 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3432 kiocb_done(kiocb, ret2, issue_flags);
3435 /* some cases will consume bytes even on error returns */
3436 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3437 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3438 return ret ?: -EAGAIN;
3441 /* it's reportedly faster than delegating the null check to kfree() */
3447 static int io_renameat_prep(struct io_kiocb *req,
3448 const struct io_uring_sqe *sqe)
3450 struct io_rename *ren = &req->rename;
3451 const char __user *oldf, *newf;
3453 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3456 ren->old_dfd = READ_ONCE(sqe->fd);
3457 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3458 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3459 ren->new_dfd = READ_ONCE(sqe->len);
3460 ren->flags = READ_ONCE(sqe->rename_flags);
3462 ren->oldpath = getname(oldf);
3463 if (IS_ERR(ren->oldpath))
3464 return PTR_ERR(ren->oldpath);
3466 ren->newpath = getname(newf);
3467 if (IS_ERR(ren->newpath)) {
3468 putname(ren->oldpath);
3469 return PTR_ERR(ren->newpath);
3472 req->flags |= REQ_F_NEED_CLEANUP;
3476 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3478 struct io_rename *ren = &req->rename;
3481 if (issue_flags & IO_URING_F_NONBLOCK)
3484 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3485 ren->newpath, ren->flags);
3487 req->flags &= ~REQ_F_NEED_CLEANUP;
3489 req_set_fail_links(req);
3490 io_req_complete(req, ret);
3494 static int io_unlinkat_prep(struct io_kiocb *req,
3495 const struct io_uring_sqe *sqe)
3497 struct io_unlink *un = &req->unlink;
3498 const char __user *fname;
3500 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3503 un->dfd = READ_ONCE(sqe->fd);
3505 un->flags = READ_ONCE(sqe->unlink_flags);
3506 if (un->flags & ~AT_REMOVEDIR)
3509 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3510 un->filename = getname(fname);
3511 if (IS_ERR(un->filename))
3512 return PTR_ERR(un->filename);
3514 req->flags |= REQ_F_NEED_CLEANUP;
3518 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3520 struct io_unlink *un = &req->unlink;
3523 if (issue_flags & IO_URING_F_NONBLOCK)
3526 if (un->flags & AT_REMOVEDIR)
3527 ret = do_rmdir(un->dfd, un->filename);
3529 ret = do_unlinkat(un->dfd, un->filename);
3531 req->flags &= ~REQ_F_NEED_CLEANUP;
3533 req_set_fail_links(req);
3534 io_req_complete(req, ret);
3538 static int io_shutdown_prep(struct io_kiocb *req,
3539 const struct io_uring_sqe *sqe)
3541 #if defined(CONFIG_NET)
3542 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3544 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3548 req->shutdown.how = READ_ONCE(sqe->len);
3555 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3557 #if defined(CONFIG_NET)
3558 struct socket *sock;
3561 if (issue_flags & IO_URING_F_NONBLOCK)
3564 sock = sock_from_file(req->file);
3565 if (unlikely(!sock))
3568 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3570 req_set_fail_links(req);
3571 io_req_complete(req, ret);
3578 static int __io_splice_prep(struct io_kiocb *req,
3579 const struct io_uring_sqe *sqe)
3581 struct io_splice* sp = &req->splice;
3582 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3584 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3588 sp->len = READ_ONCE(sqe->len);
3589 sp->flags = READ_ONCE(sqe->splice_flags);
3591 if (unlikely(sp->flags & ~valid_flags))
3594 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3595 (sp->flags & SPLICE_F_FD_IN_FIXED));
3598 req->flags |= REQ_F_NEED_CLEANUP;
3600 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3602 * Splice operation will be punted aync, and here need to
3603 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3605 req->work.flags |= IO_WQ_WORK_UNBOUND;
3611 static int io_tee_prep(struct io_kiocb *req,
3612 const struct io_uring_sqe *sqe)
3614 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3616 return __io_splice_prep(req, sqe);
3619 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3621 struct io_splice *sp = &req->splice;
3622 struct file *in = sp->file_in;
3623 struct file *out = sp->file_out;
3624 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3627 if (issue_flags & IO_URING_F_NONBLOCK)
3630 ret = do_tee(in, out, sp->len, flags);
3632 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3633 req->flags &= ~REQ_F_NEED_CLEANUP;
3636 req_set_fail_links(req);
3637 io_req_complete(req, ret);
3641 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3643 struct io_splice* sp = &req->splice;
3645 sp->off_in = READ_ONCE(sqe->splice_off_in);
3646 sp->off_out = READ_ONCE(sqe->off);
3647 return __io_splice_prep(req, sqe);
3650 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3652 struct io_splice *sp = &req->splice;
3653 struct file *in = sp->file_in;
3654 struct file *out = sp->file_out;
3655 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3656 loff_t *poff_in, *poff_out;
3659 if (issue_flags & IO_URING_F_NONBLOCK)
3662 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3663 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3666 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3668 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3669 req->flags &= ~REQ_F_NEED_CLEANUP;
3672 req_set_fail_links(req);
3673 io_req_complete(req, ret);
3678 * IORING_OP_NOP just posts a completion event, nothing else.
3680 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
3682 struct io_ring_ctx *ctx = req->ctx;
3684 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3687 __io_req_complete(req, issue_flags, 0, 0);
3691 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3693 struct io_ring_ctx *ctx = req->ctx;
3698 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3700 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3703 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3704 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3707 req->sync.off = READ_ONCE(sqe->off);
3708 req->sync.len = READ_ONCE(sqe->len);
3712 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
3714 loff_t end = req->sync.off + req->sync.len;
3717 /* fsync always requires a blocking context */
3718 if (issue_flags & IO_URING_F_NONBLOCK)
3721 ret = vfs_fsync_range(req->file, req->sync.off,
3722 end > 0 ? end : LLONG_MAX,
3723 req->sync.flags & IORING_FSYNC_DATASYNC);
3725 req_set_fail_links(req);
3726 io_req_complete(req, ret);
3730 static int io_fallocate_prep(struct io_kiocb *req,
3731 const struct io_uring_sqe *sqe)
3733 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3735 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3738 req->sync.off = READ_ONCE(sqe->off);
3739 req->sync.len = READ_ONCE(sqe->addr);
3740 req->sync.mode = READ_ONCE(sqe->len);
3744 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
3748 /* fallocate always requiring blocking context */
3749 if (issue_flags & IO_URING_F_NONBLOCK)
3751 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3754 req_set_fail_links(req);
3755 io_req_complete(req, ret);
3759 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3761 const char __user *fname;
3764 if (unlikely(sqe->ioprio || sqe->buf_index))
3766 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3769 /* open.how should be already initialised */
3770 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3771 req->open.how.flags |= O_LARGEFILE;
3773 req->open.dfd = READ_ONCE(sqe->fd);
3774 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3775 req->open.filename = getname(fname);
3776 if (IS_ERR(req->open.filename)) {
3777 ret = PTR_ERR(req->open.filename);
3778 req->open.filename = NULL;
3781 req->open.nofile = rlimit(RLIMIT_NOFILE);
3782 req->flags |= REQ_F_NEED_CLEANUP;
3786 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3790 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3792 mode = READ_ONCE(sqe->len);
3793 flags = READ_ONCE(sqe->open_flags);
3794 req->open.how = build_open_how(flags, mode);
3795 return __io_openat_prep(req, sqe);
3798 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3800 struct open_how __user *how;
3804 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3806 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3807 len = READ_ONCE(sqe->len);
3808 if (len < OPEN_HOW_SIZE_VER0)
3811 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3816 return __io_openat_prep(req, sqe);
3819 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
3821 struct open_flags op;
3824 bool resolve_nonblock;
3827 ret = build_open_flags(&req->open.how, &op);
3830 nonblock_set = op.open_flag & O_NONBLOCK;
3831 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
3832 if (issue_flags & IO_URING_F_NONBLOCK) {
3834 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3835 * it'll always -EAGAIN
3837 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
3839 op.lookup_flags |= LOOKUP_CACHED;
3840 op.open_flag |= O_NONBLOCK;
3843 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3847 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3848 /* only retry if RESOLVE_CACHED wasn't already set by application */
3849 if ((!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)) &&
3850 file == ERR_PTR(-EAGAIN)) {
3852 * We could hang on to this 'fd', but seems like marginal
3853 * gain for something that is now known to be a slower path.
3854 * So just put it, and we'll get a new one when we retry.
3862 ret = PTR_ERR(file);
3864 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
3865 file->f_flags &= ~O_NONBLOCK;
3866 fsnotify_open(file);
3867 fd_install(ret, file);
3870 putname(req->open.filename);
3871 req->flags &= ~REQ_F_NEED_CLEANUP;
3873 req_set_fail_links(req);
3874 io_req_complete(req, ret);
3878 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
3880 return io_openat2(req, issue_flags);
3883 static int io_remove_buffers_prep(struct io_kiocb *req,
3884 const struct io_uring_sqe *sqe)
3886 struct io_provide_buf *p = &req->pbuf;
3889 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3892 tmp = READ_ONCE(sqe->fd);
3893 if (!tmp || tmp > USHRT_MAX)
3896 memset(p, 0, sizeof(*p));
3898 p->bgid = READ_ONCE(sqe->buf_group);
3902 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3903 int bgid, unsigned nbufs)
3907 /* shouldn't happen */
3911 /* the head kbuf is the list itself */
3912 while (!list_empty(&buf->list)) {
3913 struct io_buffer *nxt;
3915 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3916 list_del(&nxt->list);
3923 xa_erase(&ctx->io_buffers, bgid);
3928 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
3930 struct io_provide_buf *p = &req->pbuf;
3931 struct io_ring_ctx *ctx = req->ctx;
3932 struct io_buffer *head;
3934 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3936 io_ring_submit_lock(ctx, !force_nonblock);
3938 lockdep_assert_held(&ctx->uring_lock);
3941 head = xa_load(&ctx->io_buffers, p->bgid);
3943 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3945 req_set_fail_links(req);
3947 /* need to hold the lock to complete IOPOLL requests */
3948 if (ctx->flags & IORING_SETUP_IOPOLL) {
3949 __io_req_complete(req, issue_flags, ret, 0);
3950 io_ring_submit_unlock(ctx, !force_nonblock);
3952 io_ring_submit_unlock(ctx, !force_nonblock);
3953 __io_req_complete(req, issue_flags, ret, 0);
3958 static int io_provide_buffers_prep(struct io_kiocb *req,
3959 const struct io_uring_sqe *sqe)
3962 struct io_provide_buf *p = &req->pbuf;
3965 if (sqe->ioprio || sqe->rw_flags)
3968 tmp = READ_ONCE(sqe->fd);
3969 if (!tmp || tmp > USHRT_MAX)
3972 p->addr = READ_ONCE(sqe->addr);
3973 p->len = READ_ONCE(sqe->len);
3975 size = (unsigned long)p->len * p->nbufs;
3976 if (!access_ok(u64_to_user_ptr(p->addr), size))
3979 p->bgid = READ_ONCE(sqe->buf_group);
3980 tmp = READ_ONCE(sqe->off);
3981 if (tmp > USHRT_MAX)
3987 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3989 struct io_buffer *buf;
3990 u64 addr = pbuf->addr;
3991 int i, bid = pbuf->bid;
3993 for (i = 0; i < pbuf->nbufs; i++) {
3994 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3999 buf->len = pbuf->len;
4004 INIT_LIST_HEAD(&buf->list);
4007 list_add_tail(&buf->list, &(*head)->list);
4011 return i ? i : -ENOMEM;
4014 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4016 struct io_provide_buf *p = &req->pbuf;
4017 struct io_ring_ctx *ctx = req->ctx;
4018 struct io_buffer *head, *list;
4020 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4022 io_ring_submit_lock(ctx, !force_nonblock);
4024 lockdep_assert_held(&ctx->uring_lock);
4026 list = head = xa_load(&ctx->io_buffers, p->bgid);
4028 ret = io_add_buffers(p, &head);
4029 if (ret >= 0 && !list) {
4030 ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4032 __io_remove_buffers(ctx, head, p->bgid, -1U);
4035 req_set_fail_links(req);
4037 /* need to hold the lock to complete IOPOLL requests */
4038 if (ctx->flags & IORING_SETUP_IOPOLL) {
4039 __io_req_complete(req, issue_flags, ret, 0);
4040 io_ring_submit_unlock(ctx, !force_nonblock);
4042 io_ring_submit_unlock(ctx, !force_nonblock);
4043 __io_req_complete(req, issue_flags, ret, 0);
4048 static int io_epoll_ctl_prep(struct io_kiocb *req,
4049 const struct io_uring_sqe *sqe)
4051 #if defined(CONFIG_EPOLL)
4052 if (sqe->ioprio || sqe->buf_index)
4054 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4057 req->epoll.epfd = READ_ONCE(sqe->fd);
4058 req->epoll.op = READ_ONCE(sqe->len);
4059 req->epoll.fd = READ_ONCE(sqe->off);
4061 if (ep_op_has_event(req->epoll.op)) {
4062 struct epoll_event __user *ev;
4064 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4065 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4075 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4077 #if defined(CONFIG_EPOLL)
4078 struct io_epoll *ie = &req->epoll;
4080 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4082 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4083 if (force_nonblock && ret == -EAGAIN)
4087 req_set_fail_links(req);
4088 __io_req_complete(req, issue_flags, ret, 0);
4095 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4097 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4098 if (sqe->ioprio || sqe->buf_index || sqe->off)
4100 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4103 req->madvise.addr = READ_ONCE(sqe->addr);
4104 req->madvise.len = READ_ONCE(sqe->len);
4105 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4112 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4114 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4115 struct io_madvise *ma = &req->madvise;
4118 if (issue_flags & IO_URING_F_NONBLOCK)
4121 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4123 req_set_fail_links(req);
4124 io_req_complete(req, ret);
4131 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4133 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4135 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4138 req->fadvise.offset = READ_ONCE(sqe->off);
4139 req->fadvise.len = READ_ONCE(sqe->len);
4140 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4144 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4146 struct io_fadvise *fa = &req->fadvise;
4149 if (issue_flags & IO_URING_F_NONBLOCK) {
4150 switch (fa->advice) {
4151 case POSIX_FADV_NORMAL:
4152 case POSIX_FADV_RANDOM:
4153 case POSIX_FADV_SEQUENTIAL:
4160 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4162 req_set_fail_links(req);
4163 io_req_complete(req, ret);
4167 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4169 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4171 if (sqe->ioprio || sqe->buf_index)
4173 if (req->flags & REQ_F_FIXED_FILE)
4176 req->statx.dfd = READ_ONCE(sqe->fd);
4177 req->statx.mask = READ_ONCE(sqe->len);
4178 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4179 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4180 req->statx.flags = READ_ONCE(sqe->statx_flags);
4185 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4187 struct io_statx *ctx = &req->statx;
4190 if (issue_flags & IO_URING_F_NONBLOCK) {
4191 /* only need file table for an actual valid fd */
4192 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
4193 req->flags |= REQ_F_NO_FILE_TABLE;
4197 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4201 req_set_fail_links(req);
4202 io_req_complete(req, ret);
4206 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4208 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4210 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4211 sqe->rw_flags || sqe->buf_index)
4213 if (req->flags & REQ_F_FIXED_FILE)
4216 req->close.fd = READ_ONCE(sqe->fd);
4220 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4222 struct files_struct *files = current->files;
4223 struct io_close *close = &req->close;
4224 struct fdtable *fdt;
4230 spin_lock(&files->file_lock);
4231 fdt = files_fdtable(files);
4232 if (close->fd >= fdt->max_fds) {
4233 spin_unlock(&files->file_lock);
4236 file = fdt->fd[close->fd];
4238 spin_unlock(&files->file_lock);
4242 if (file->f_op == &io_uring_fops) {
4243 spin_unlock(&files->file_lock);
4248 /* if the file has a flush method, be safe and punt to async */
4249 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4250 spin_unlock(&files->file_lock);
4254 ret = __close_fd_get_file(close->fd, &file);
4255 spin_unlock(&files->file_lock);
4262 /* No ->flush() or already async, safely close from here */
4263 ret = filp_close(file, current->files);
4266 req_set_fail_links(req);
4269 __io_req_complete(req, issue_flags, ret, 0);
4273 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4275 struct io_ring_ctx *ctx = req->ctx;
4277 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4279 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4282 req->sync.off = READ_ONCE(sqe->off);
4283 req->sync.len = READ_ONCE(sqe->len);
4284 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4288 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4292 /* sync_file_range always requires a blocking context */
4293 if (issue_flags & IO_URING_F_NONBLOCK)
4296 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4299 req_set_fail_links(req);
4300 io_req_complete(req, ret);
4304 #if defined(CONFIG_NET)
4305 static int io_setup_async_msg(struct io_kiocb *req,
4306 struct io_async_msghdr *kmsg)
4308 struct io_async_msghdr *async_msg = req->async_data;
4312 if (io_alloc_async_data(req)) {
4313 kfree(kmsg->free_iov);
4316 async_msg = req->async_data;
4317 req->flags |= REQ_F_NEED_CLEANUP;
4318 memcpy(async_msg, kmsg, sizeof(*kmsg));
4319 async_msg->msg.msg_name = &async_msg->addr;
4320 /* if were using fast_iov, set it to the new one */
4321 if (!async_msg->free_iov)
4322 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4327 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4328 struct io_async_msghdr *iomsg)
4330 iomsg->msg.msg_name = &iomsg->addr;
4331 iomsg->free_iov = iomsg->fast_iov;
4332 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4333 req->sr_msg.msg_flags, &iomsg->free_iov);
4336 static int io_sendmsg_prep_async(struct io_kiocb *req)
4340 if (!io_op_defs[req->opcode].needs_async_data)
4342 ret = io_sendmsg_copy_hdr(req, req->async_data);
4344 req->flags |= REQ_F_NEED_CLEANUP;
4348 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4350 struct io_sr_msg *sr = &req->sr_msg;
4352 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4355 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4356 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4357 sr->len = READ_ONCE(sqe->len);
4359 #ifdef CONFIG_COMPAT
4360 if (req->ctx->compat)
4361 sr->msg_flags |= MSG_CMSG_COMPAT;
4366 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4368 struct io_async_msghdr iomsg, *kmsg;
4369 struct socket *sock;
4374 sock = sock_from_file(req->file);
4375 if (unlikely(!sock))
4378 kmsg = req->async_data;
4380 ret = io_sendmsg_copy_hdr(req, &iomsg);
4386 flags = req->sr_msg.msg_flags | MSG_NOSIGNAL;
4387 if (flags & MSG_DONTWAIT)
4388 req->flags |= REQ_F_NOWAIT;
4389 else if (issue_flags & IO_URING_F_NONBLOCK)
4390 flags |= MSG_DONTWAIT;
4392 if (flags & MSG_WAITALL)
4393 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4395 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4396 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4397 return io_setup_async_msg(req, kmsg);
4398 if (ret == -ERESTARTSYS)
4401 /* fast path, check for non-NULL to avoid function call */
4403 kfree(kmsg->free_iov);
4404 req->flags &= ~REQ_F_NEED_CLEANUP;
4406 req_set_fail_links(req);
4407 __io_req_complete(req, issue_flags, ret, 0);
4411 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4413 struct io_sr_msg *sr = &req->sr_msg;
4416 struct socket *sock;
4421 sock = sock_from_file(req->file);
4422 if (unlikely(!sock))
4425 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4429 msg.msg_name = NULL;
4430 msg.msg_control = NULL;
4431 msg.msg_controllen = 0;
4432 msg.msg_namelen = 0;
4434 flags = req->sr_msg.msg_flags | MSG_NOSIGNAL;
4435 if (flags & MSG_DONTWAIT)
4436 req->flags |= REQ_F_NOWAIT;
4437 else if (issue_flags & IO_URING_F_NONBLOCK)
4438 flags |= MSG_DONTWAIT;
4440 if (flags & MSG_WAITALL)
4441 min_ret = iov_iter_count(&msg.msg_iter);
4443 msg.msg_flags = flags;
4444 ret = sock_sendmsg(sock, &msg);
4445 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4447 if (ret == -ERESTARTSYS)
4451 req_set_fail_links(req);
4452 __io_req_complete(req, issue_flags, ret, 0);
4456 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4457 struct io_async_msghdr *iomsg)
4459 struct io_sr_msg *sr = &req->sr_msg;
4460 struct iovec __user *uiov;
4464 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4465 &iomsg->uaddr, &uiov, &iov_len);
4469 if (req->flags & REQ_F_BUFFER_SELECT) {
4472 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4474 sr->len = iomsg->fast_iov[0].iov_len;
4475 iomsg->free_iov = NULL;
4477 iomsg->free_iov = iomsg->fast_iov;
4478 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4479 &iomsg->free_iov, &iomsg->msg.msg_iter,
4488 #ifdef CONFIG_COMPAT
4489 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4490 struct io_async_msghdr *iomsg)
4492 struct compat_msghdr __user *msg_compat;
4493 struct io_sr_msg *sr = &req->sr_msg;
4494 struct compat_iovec __user *uiov;
4499 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4500 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4505 uiov = compat_ptr(ptr);
4506 if (req->flags & REQ_F_BUFFER_SELECT) {
4507 compat_ssize_t clen;
4511 if (!access_ok(uiov, sizeof(*uiov)))
4513 if (__get_user(clen, &uiov->iov_len))
4518 iomsg->free_iov = NULL;
4520 iomsg->free_iov = iomsg->fast_iov;
4521 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4522 UIO_FASTIOV, &iomsg->free_iov,
4523 &iomsg->msg.msg_iter, true);
4532 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4533 struct io_async_msghdr *iomsg)
4535 iomsg->msg.msg_name = &iomsg->addr;
4537 #ifdef CONFIG_COMPAT
4538 if (req->ctx->compat)
4539 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4542 return __io_recvmsg_copy_hdr(req, iomsg);
4545 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4548 struct io_sr_msg *sr = &req->sr_msg;
4549 struct io_buffer *kbuf;
4551 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4556 req->flags |= REQ_F_BUFFER_SELECTED;
4560 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4562 return io_put_kbuf(req, req->sr_msg.kbuf);
4565 static int io_recvmsg_prep_async(struct io_kiocb *req)
4569 if (!io_op_defs[req->opcode].needs_async_data)
4571 ret = io_recvmsg_copy_hdr(req, req->async_data);
4573 req->flags |= REQ_F_NEED_CLEANUP;
4577 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4579 struct io_sr_msg *sr = &req->sr_msg;
4581 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4584 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4585 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4586 sr->len = READ_ONCE(sqe->len);
4587 sr->bgid = READ_ONCE(sqe->buf_group);
4589 #ifdef CONFIG_COMPAT
4590 if (req->ctx->compat)
4591 sr->msg_flags |= MSG_CMSG_COMPAT;
4596 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4598 struct io_async_msghdr iomsg, *kmsg;
4599 struct socket *sock;
4600 struct io_buffer *kbuf;
4603 int ret, cflags = 0;
4604 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4606 sock = sock_from_file(req->file);
4607 if (unlikely(!sock))
4610 kmsg = req->async_data;
4612 ret = io_recvmsg_copy_hdr(req, &iomsg);
4618 if (req->flags & REQ_F_BUFFER_SELECT) {
4619 kbuf = io_recv_buffer_select(req, !force_nonblock);
4621 return PTR_ERR(kbuf);
4622 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4623 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4624 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4625 1, req->sr_msg.len);
4628 flags = req->sr_msg.msg_flags | MSG_NOSIGNAL;
4629 if (flags & MSG_DONTWAIT)
4630 req->flags |= REQ_F_NOWAIT;
4631 else if (force_nonblock)
4632 flags |= MSG_DONTWAIT;
4634 if (flags & MSG_WAITALL)
4635 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4637 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4638 kmsg->uaddr, flags);
4639 if (force_nonblock && ret == -EAGAIN)
4640 return io_setup_async_msg(req, kmsg);
4641 if (ret == -ERESTARTSYS)
4644 if (req->flags & REQ_F_BUFFER_SELECTED)
4645 cflags = io_put_recv_kbuf(req);
4646 /* fast path, check for non-NULL to avoid function call */
4648 kfree(kmsg->free_iov);
4649 req->flags &= ~REQ_F_NEED_CLEANUP;
4650 if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4651 req_set_fail_links(req);
4652 __io_req_complete(req, issue_flags, ret, cflags);
4656 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4658 struct io_buffer *kbuf;
4659 struct io_sr_msg *sr = &req->sr_msg;
4661 void __user *buf = sr->buf;
4662 struct socket *sock;
4666 int ret, cflags = 0;
4667 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4669 sock = sock_from_file(req->file);
4670 if (unlikely(!sock))
4673 if (req->flags & REQ_F_BUFFER_SELECT) {
4674 kbuf = io_recv_buffer_select(req, !force_nonblock);
4676 return PTR_ERR(kbuf);
4677 buf = u64_to_user_ptr(kbuf->addr);
4680 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4684 msg.msg_name = NULL;
4685 msg.msg_control = NULL;
4686 msg.msg_controllen = 0;
4687 msg.msg_namelen = 0;
4688 msg.msg_iocb = NULL;
4691 flags = req->sr_msg.msg_flags | MSG_NOSIGNAL;
4692 if (flags & MSG_DONTWAIT)
4693 req->flags |= REQ_F_NOWAIT;
4694 else if (force_nonblock)
4695 flags |= MSG_DONTWAIT;
4697 if (flags & MSG_WAITALL)
4698 min_ret = iov_iter_count(&msg.msg_iter);
4700 ret = sock_recvmsg(sock, &msg, flags);
4701 if (force_nonblock && ret == -EAGAIN)
4703 if (ret == -ERESTARTSYS)
4706 if (req->flags & REQ_F_BUFFER_SELECTED)
4707 cflags = io_put_recv_kbuf(req);
4708 if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4709 req_set_fail_links(req);
4710 __io_req_complete(req, issue_flags, ret, cflags);
4714 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4716 struct io_accept *accept = &req->accept;
4718 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4720 if (sqe->ioprio || sqe->len || sqe->buf_index)
4723 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4724 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4725 accept->flags = READ_ONCE(sqe->accept_flags);
4726 accept->nofile = rlimit(RLIMIT_NOFILE);
4730 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
4732 struct io_accept *accept = &req->accept;
4733 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4734 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4737 if (req->file->f_flags & O_NONBLOCK)
4738 req->flags |= REQ_F_NOWAIT;
4740 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4741 accept->addr_len, accept->flags,
4743 if (ret == -EAGAIN && force_nonblock)
4746 if (ret == -ERESTARTSYS)
4748 req_set_fail_links(req);
4750 __io_req_complete(req, issue_flags, ret, 0);
4754 static int io_connect_prep_async(struct io_kiocb *req)
4756 struct io_async_connect *io = req->async_data;
4757 struct io_connect *conn = &req->connect;
4759 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4762 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4764 struct io_connect *conn = &req->connect;
4766 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4768 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4771 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4772 conn->addr_len = READ_ONCE(sqe->addr2);
4776 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
4778 struct io_async_connect __io, *io;
4779 unsigned file_flags;
4781 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4783 if (req->async_data) {
4784 io = req->async_data;
4786 ret = move_addr_to_kernel(req->connect.addr,
4787 req->connect.addr_len,
4794 file_flags = force_nonblock ? O_NONBLOCK : 0;
4796 ret = __sys_connect_file(req->file, &io->address,
4797 req->connect.addr_len, file_flags);
4798 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4799 if (req->async_data)
4801 if (io_alloc_async_data(req)) {
4805 memcpy(req->async_data, &__io, sizeof(__io));
4808 if (ret == -ERESTARTSYS)
4812 req_set_fail_links(req);
4813 __io_req_complete(req, issue_flags, ret, 0);
4816 #else /* !CONFIG_NET */
4817 #define IO_NETOP_FN(op) \
4818 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
4820 return -EOPNOTSUPP; \
4823 #define IO_NETOP_PREP(op) \
4825 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4827 return -EOPNOTSUPP; \
4830 #define IO_NETOP_PREP_ASYNC(op) \
4832 static int io_##op##_prep_async(struct io_kiocb *req) \
4834 return -EOPNOTSUPP; \
4837 IO_NETOP_PREP_ASYNC(sendmsg);
4838 IO_NETOP_PREP_ASYNC(recvmsg);
4839 IO_NETOP_PREP_ASYNC(connect);
4840 IO_NETOP_PREP(accept);
4843 #endif /* CONFIG_NET */
4845 struct io_poll_table {
4846 struct poll_table_struct pt;
4847 struct io_kiocb *req;
4851 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4852 __poll_t mask, task_work_func_t func)
4856 /* for instances that support it check for an event match first: */
4857 if (mask && !(mask & poll->events))
4860 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4862 list_del_init(&poll->wait.entry);
4865 req->task_work.func = func;
4866 percpu_ref_get(&req->ctx->refs);
4869 * If this fails, then the task is exiting. When a task exits, the
4870 * work gets canceled, so just cancel this request as well instead
4871 * of executing it. We can't safely execute it anyway, as we may not
4872 * have the needed state needed for it anyway.
4874 ret = io_req_task_work_add(req);
4875 if (unlikely(ret)) {
4876 WRITE_ONCE(poll->canceled, true);
4877 io_req_task_work_add_fallback(req, func);
4882 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4883 __acquires(&req->ctx->completion_lock)
4885 struct io_ring_ctx *ctx = req->ctx;
4887 if (!req->result && !READ_ONCE(poll->canceled)) {
4888 struct poll_table_struct pt = { ._key = poll->events };
4890 req->result = vfs_poll(req->file, &pt) & poll->events;
4893 spin_lock_irq(&ctx->completion_lock);
4894 if (!req->result && !READ_ONCE(poll->canceled)) {
4895 add_wait_queue(poll->head, &poll->wait);
4902 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4904 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4905 if (req->opcode == IORING_OP_POLL_ADD)
4906 return req->async_data;
4907 return req->apoll->double_poll;
4910 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4912 if (req->opcode == IORING_OP_POLL_ADD)
4914 return &req->apoll->poll;
4917 static void io_poll_remove_double(struct io_kiocb *req)
4919 struct io_poll_iocb *poll = io_poll_get_double(req);
4921 lockdep_assert_held(&req->ctx->completion_lock);
4923 if (poll && poll->head) {
4924 struct wait_queue_head *head = poll->head;
4926 spin_lock(&head->lock);
4927 list_del_init(&poll->wait.entry);
4928 if (poll->wait.private)
4929 refcount_dec(&req->refs);
4931 spin_unlock(&head->lock);
4935 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4937 struct io_ring_ctx *ctx = req->ctx;
4939 io_poll_remove_double(req);
4940 req->poll.done = true;
4941 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4942 io_commit_cqring(ctx);
4945 static void io_poll_task_func(struct callback_head *cb)
4947 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4948 struct io_ring_ctx *ctx = req->ctx;
4949 struct io_kiocb *nxt;
4951 if (io_poll_rewait(req, &req->poll)) {
4952 spin_unlock_irq(&ctx->completion_lock);
4954 hash_del(&req->hash_node);
4955 io_poll_complete(req, req->result, 0);
4956 spin_unlock_irq(&ctx->completion_lock);
4958 nxt = io_put_req_find_next(req);
4959 io_cqring_ev_posted(ctx);
4961 __io_req_task_submit(nxt);
4964 percpu_ref_put(&ctx->refs);
4967 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4968 int sync, void *key)
4970 struct io_kiocb *req = wait->private;
4971 struct io_poll_iocb *poll = io_poll_get_single(req);
4972 __poll_t mask = key_to_poll(key);
4974 /* for instances that support it check for an event match first: */
4975 if (mask && !(mask & poll->events))
4978 list_del_init(&wait->entry);
4980 if (poll && poll->head) {
4983 spin_lock(&poll->head->lock);
4984 done = list_empty(&poll->wait.entry);
4986 list_del_init(&poll->wait.entry);
4987 /* make sure double remove sees this as being gone */
4988 wait->private = NULL;
4989 spin_unlock(&poll->head->lock);
4991 /* use wait func handler, so it matches the rq type */
4992 poll->wait.func(&poll->wait, mode, sync, key);
4995 refcount_dec(&req->refs);
4999 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5000 wait_queue_func_t wake_func)
5004 poll->canceled = false;
5005 poll->events = events;
5006 INIT_LIST_HEAD(&poll->wait.entry);
5007 init_waitqueue_func_entry(&poll->wait, wake_func);
5010 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5011 struct wait_queue_head *head,
5012 struct io_poll_iocb **poll_ptr)
5014 struct io_kiocb *req = pt->req;
5017 * If poll->head is already set, it's because the file being polled
5018 * uses multiple waitqueues for poll handling (eg one for read, one
5019 * for write). Setup a separate io_poll_iocb if this happens.
5021 if (unlikely(poll->head)) {
5022 struct io_poll_iocb *poll_one = poll;
5024 /* already have a 2nd entry, fail a third attempt */
5026 pt->error = -EINVAL;
5029 /* double add on the same waitqueue head, ignore */
5030 if (poll->head == head)
5032 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5034 pt->error = -ENOMEM;
5037 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5038 refcount_inc(&req->refs);
5039 poll->wait.private = req;
5046 if (poll->events & EPOLLEXCLUSIVE)
5047 add_wait_queue_exclusive(head, &poll->wait);
5049 add_wait_queue(head, &poll->wait);
5052 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5053 struct poll_table_struct *p)
5055 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5056 struct async_poll *apoll = pt->req->apoll;
5058 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5061 static void io_async_task_func(struct callback_head *cb)
5063 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5064 struct async_poll *apoll = req->apoll;
5065 struct io_ring_ctx *ctx = req->ctx;
5067 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
5069 if (io_poll_rewait(req, &apoll->poll)) {
5070 spin_unlock_irq(&ctx->completion_lock);
5071 percpu_ref_put(&ctx->refs);
5075 /* If req is still hashed, it cannot have been canceled. Don't check. */
5076 if (hash_hashed(&req->hash_node))
5077 hash_del(&req->hash_node);
5079 io_poll_remove_double(req);
5080 spin_unlock_irq(&ctx->completion_lock);
5082 if (!READ_ONCE(apoll->poll.canceled))
5083 __io_req_task_submit(req);
5085 __io_req_task_cancel(req, -ECANCELED);
5087 percpu_ref_put(&ctx->refs);
5088 kfree(apoll->double_poll);
5092 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5095 struct io_kiocb *req = wait->private;
5096 struct io_poll_iocb *poll = &req->apoll->poll;
5098 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5101 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5104 static void io_poll_req_insert(struct io_kiocb *req)
5106 struct io_ring_ctx *ctx = req->ctx;
5107 struct hlist_head *list;
5109 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5110 hlist_add_head(&req->hash_node, list);
5113 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5114 struct io_poll_iocb *poll,
5115 struct io_poll_table *ipt, __poll_t mask,
5116 wait_queue_func_t wake_func)
5117 __acquires(&ctx->completion_lock)
5119 struct io_ring_ctx *ctx = req->ctx;
5120 bool cancel = false;
5122 INIT_HLIST_NODE(&req->hash_node);
5123 io_init_poll_iocb(poll, mask, wake_func);
5124 poll->file = req->file;
5125 poll->wait.private = req;
5127 ipt->pt._key = mask;
5129 ipt->error = -EINVAL;
5131 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5133 spin_lock_irq(&ctx->completion_lock);
5134 if (likely(poll->head)) {
5135 spin_lock(&poll->head->lock);
5136 if (unlikely(list_empty(&poll->wait.entry))) {
5142 if (mask || ipt->error)
5143 list_del_init(&poll->wait.entry);
5145 WRITE_ONCE(poll->canceled, true);
5146 else if (!poll->done) /* actually waiting for an event */
5147 io_poll_req_insert(req);
5148 spin_unlock(&poll->head->lock);
5154 static bool io_arm_poll_handler(struct io_kiocb *req)
5156 const struct io_op_def *def = &io_op_defs[req->opcode];
5157 struct io_ring_ctx *ctx = req->ctx;
5158 struct async_poll *apoll;
5159 struct io_poll_table ipt;
5163 if (!req->file || !file_can_poll(req->file))
5165 if (req->flags & REQ_F_POLLED)
5169 else if (def->pollout)
5173 /* if we can't nonblock try, then no point in arming a poll handler */
5174 if (!io_file_supports_async(req->file, rw))
5177 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5178 if (unlikely(!apoll))
5180 apoll->double_poll = NULL;
5182 req->flags |= REQ_F_POLLED;
5187 mask |= POLLIN | POLLRDNORM;
5189 mask |= POLLOUT | POLLWRNORM;
5191 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5192 if ((req->opcode == IORING_OP_RECVMSG) &&
5193 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5196 mask |= POLLERR | POLLPRI;
5198 ipt.pt._qproc = io_async_queue_proc;
5200 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5202 if (ret || ipt.error) {
5203 io_poll_remove_double(req);
5204 spin_unlock_irq(&ctx->completion_lock);
5205 kfree(apoll->double_poll);
5209 spin_unlock_irq(&ctx->completion_lock);
5210 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5211 apoll->poll.events);
5215 static bool __io_poll_remove_one(struct io_kiocb *req,
5216 struct io_poll_iocb *poll)
5218 bool do_complete = false;
5220 spin_lock(&poll->head->lock);
5221 WRITE_ONCE(poll->canceled, true);
5222 if (!list_empty(&poll->wait.entry)) {
5223 list_del_init(&poll->wait.entry);
5226 spin_unlock(&poll->head->lock);
5227 hash_del(&req->hash_node);
5231 static bool io_poll_remove_one(struct io_kiocb *req)
5235 io_poll_remove_double(req);
5237 if (req->opcode == IORING_OP_POLL_ADD) {
5238 do_complete = __io_poll_remove_one(req, &req->poll);
5240 struct async_poll *apoll = req->apoll;
5242 /* non-poll requests have submit ref still */
5243 do_complete = __io_poll_remove_one(req, &apoll->poll);
5246 kfree(apoll->double_poll);
5252 io_cqring_fill_event(req, -ECANCELED);
5253 io_commit_cqring(req->ctx);
5254 req_set_fail_links(req);
5255 io_put_req_deferred(req, 1);
5262 * Returns true if we found and killed one or more poll requests
5264 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5265 struct files_struct *files)
5267 struct hlist_node *tmp;
5268 struct io_kiocb *req;
5271 spin_lock_irq(&ctx->completion_lock);
5272 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5273 struct hlist_head *list;
5275 list = &ctx->cancel_hash[i];
5276 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5277 if (io_match_task(req, tsk, files))
5278 posted += io_poll_remove_one(req);
5281 spin_unlock_irq(&ctx->completion_lock);
5284 io_cqring_ev_posted(ctx);
5289 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5291 struct hlist_head *list;
5292 struct io_kiocb *req;
5294 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5295 hlist_for_each_entry(req, list, hash_node) {
5296 if (sqe_addr != req->user_data)
5298 if (io_poll_remove_one(req))
5306 static int io_poll_remove_prep(struct io_kiocb *req,
5307 const struct io_uring_sqe *sqe)
5309 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5311 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5315 req->poll_remove.addr = READ_ONCE(sqe->addr);
5320 * Find a running poll command that matches one specified in sqe->addr,
5321 * and remove it if found.
5323 static int io_poll_remove(struct io_kiocb *req, unsigned int issue_flags)
5325 struct io_ring_ctx *ctx = req->ctx;
5328 spin_lock_irq(&ctx->completion_lock);
5329 ret = io_poll_cancel(ctx, req->poll_remove.addr);
5330 spin_unlock_irq(&ctx->completion_lock);
5333 req_set_fail_links(req);
5334 io_req_complete(req, ret);
5338 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5341 struct io_kiocb *req = wait->private;
5342 struct io_poll_iocb *poll = &req->poll;
5344 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5347 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5348 struct poll_table_struct *p)
5350 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5352 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5355 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5357 struct io_poll_iocb *poll = &req->poll;
5360 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5362 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5365 events = READ_ONCE(sqe->poll32_events);
5367 events = swahw32(events);
5369 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5370 (events & EPOLLEXCLUSIVE);
5374 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5376 struct io_poll_iocb *poll = &req->poll;
5377 struct io_ring_ctx *ctx = req->ctx;
5378 struct io_poll_table ipt;
5381 ipt.pt._qproc = io_poll_queue_proc;
5383 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5386 if (mask) { /* no async, we'd stolen it */
5388 io_poll_complete(req, mask, 0);
5390 spin_unlock_irq(&ctx->completion_lock);
5393 io_cqring_ev_posted(ctx);
5399 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5401 struct io_timeout_data *data = container_of(timer,
5402 struct io_timeout_data, timer);
5403 struct io_kiocb *req = data->req;
5404 struct io_ring_ctx *ctx = req->ctx;
5405 unsigned long flags;
5407 spin_lock_irqsave(&ctx->completion_lock, flags);
5408 list_del_init(&req->timeout.list);
5409 atomic_set(&req->ctx->cq_timeouts,
5410 atomic_read(&req->ctx->cq_timeouts) + 1);
5412 io_cqring_fill_event(req, -ETIME);
5413 io_commit_cqring(ctx);
5414 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5416 io_cqring_ev_posted(ctx);
5417 req_set_fail_links(req);
5419 return HRTIMER_NORESTART;
5422 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5425 struct io_timeout_data *io;
5426 struct io_kiocb *req;
5429 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5430 if (user_data == req->user_data) {
5437 return ERR_PTR(ret);
5439 io = req->async_data;
5440 ret = hrtimer_try_to_cancel(&io->timer);
5442 return ERR_PTR(-EALREADY);
5443 list_del_init(&req->timeout.list);
5447 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5449 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5452 return PTR_ERR(req);
5454 req_set_fail_links(req);
5455 io_cqring_fill_event(req, -ECANCELED);
5456 io_put_req_deferred(req, 1);
5460 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5461 struct timespec64 *ts, enum hrtimer_mode mode)
5463 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5464 struct io_timeout_data *data;
5467 return PTR_ERR(req);
5469 req->timeout.off = 0; /* noseq */
5470 data = req->async_data;
5471 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5472 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5473 data->timer.function = io_timeout_fn;
5474 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5478 static int io_timeout_remove_prep(struct io_kiocb *req,
5479 const struct io_uring_sqe *sqe)
5481 struct io_timeout_rem *tr = &req->timeout_rem;
5483 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5485 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5487 if (sqe->ioprio || sqe->buf_index || sqe->len)
5490 tr->addr = READ_ONCE(sqe->addr);
5491 tr->flags = READ_ONCE(sqe->timeout_flags);
5492 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5493 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5495 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5497 } else if (tr->flags) {
5498 /* timeout removal doesn't support flags */
5505 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5507 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5512 * Remove or update an existing timeout command
5514 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5516 struct io_timeout_rem *tr = &req->timeout_rem;
5517 struct io_ring_ctx *ctx = req->ctx;
5520 spin_lock_irq(&ctx->completion_lock);
5521 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE))
5522 ret = io_timeout_cancel(ctx, tr->addr);
5524 ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5525 io_translate_timeout_mode(tr->flags));
5527 io_cqring_fill_event(req, ret);
5528 io_commit_cqring(ctx);
5529 spin_unlock_irq(&ctx->completion_lock);
5530 io_cqring_ev_posted(ctx);
5532 req_set_fail_links(req);
5537 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5538 bool is_timeout_link)
5540 struct io_timeout_data *data;
5542 u32 off = READ_ONCE(sqe->off);
5544 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5546 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5548 if (off && is_timeout_link)
5550 flags = READ_ONCE(sqe->timeout_flags);
5551 if (flags & ~IORING_TIMEOUT_ABS)
5554 req->timeout.off = off;
5556 if (!req->async_data && io_alloc_async_data(req))
5559 data = req->async_data;
5562 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5565 data->mode = io_translate_timeout_mode(flags);
5566 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5567 if (is_timeout_link)
5568 io_req_track_inflight(req);
5572 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5574 struct io_ring_ctx *ctx = req->ctx;
5575 struct io_timeout_data *data = req->async_data;
5576 struct list_head *entry;
5577 u32 tail, off = req->timeout.off;
5579 spin_lock_irq(&ctx->completion_lock);
5582 * sqe->off holds how many events that need to occur for this
5583 * timeout event to be satisfied. If it isn't set, then this is
5584 * a pure timeout request, sequence isn't used.
5586 if (io_is_timeout_noseq(req)) {
5587 entry = ctx->timeout_list.prev;
5591 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5592 req->timeout.target_seq = tail + off;
5594 /* Update the last seq here in case io_flush_timeouts() hasn't.
5595 * This is safe because ->completion_lock is held, and submissions
5596 * and completions are never mixed in the same ->completion_lock section.
5598 ctx->cq_last_tm_flush = tail;
5601 * Insertion sort, ensuring the first entry in the list is always
5602 * the one we need first.
5604 list_for_each_prev(entry, &ctx->timeout_list) {
5605 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5608 if (io_is_timeout_noseq(nxt))
5610 /* nxt.seq is behind @tail, otherwise would've been completed */
5611 if (off >= nxt->timeout.target_seq - tail)
5615 list_add(&req->timeout.list, entry);
5616 data->timer.function = io_timeout_fn;
5617 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5618 spin_unlock_irq(&ctx->completion_lock);
5622 struct io_cancel_data {
5623 struct io_ring_ctx *ctx;
5627 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5629 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5630 struct io_cancel_data *cd = data;
5632 return req->ctx == cd->ctx && req->user_data == cd->user_data;
5635 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
5636 struct io_ring_ctx *ctx)
5638 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
5639 enum io_wq_cancel cancel_ret;
5642 if (!tctx || !tctx->io_wq)
5645 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
5646 switch (cancel_ret) {
5647 case IO_WQ_CANCEL_OK:
5650 case IO_WQ_CANCEL_RUNNING:
5653 case IO_WQ_CANCEL_NOTFOUND:
5661 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5662 struct io_kiocb *req, __u64 sqe_addr,
5665 unsigned long flags;
5668 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5669 if (ret != -ENOENT) {
5670 spin_lock_irqsave(&ctx->completion_lock, flags);
5674 spin_lock_irqsave(&ctx->completion_lock, flags);
5675 ret = io_timeout_cancel(ctx, sqe_addr);
5678 ret = io_poll_cancel(ctx, sqe_addr);
5682 io_cqring_fill_event(req, ret);
5683 io_commit_cqring(ctx);
5684 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5685 io_cqring_ev_posted(ctx);
5688 req_set_fail_links(req);
5692 static int io_async_cancel_prep(struct io_kiocb *req,
5693 const struct io_uring_sqe *sqe)
5695 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5697 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5699 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5702 req->cancel.addr = READ_ONCE(sqe->addr);
5706 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
5708 struct io_ring_ctx *ctx = req->ctx;
5709 u64 sqe_addr = req->cancel.addr;
5710 struct io_tctx_node *node;
5713 /* tasks should wait for their io-wq threads, so safe w/o sync */
5714 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5715 spin_lock_irq(&ctx->completion_lock);
5718 ret = io_timeout_cancel(ctx, sqe_addr);
5721 ret = io_poll_cancel(ctx, sqe_addr);
5724 spin_unlock_irq(&ctx->completion_lock);
5726 /* slow path, try all io-wq's */
5727 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5729 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
5730 struct io_uring_task *tctx = node->task->io_uring;
5732 if (!tctx || !tctx->io_wq)
5734 ret = io_async_cancel_one(tctx, req->cancel.addr, ctx);
5738 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5740 spin_lock_irq(&ctx->completion_lock);
5742 io_cqring_fill_event(req, ret);
5743 io_commit_cqring(ctx);
5744 spin_unlock_irq(&ctx->completion_lock);
5745 io_cqring_ev_posted(ctx);
5748 req_set_fail_links(req);
5753 static int io_rsrc_update_prep(struct io_kiocb *req,
5754 const struct io_uring_sqe *sqe)
5756 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
5758 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5760 if (sqe->ioprio || sqe->rw_flags)
5763 req->rsrc_update.offset = READ_ONCE(sqe->off);
5764 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
5765 if (!req->rsrc_update.nr_args)
5767 req->rsrc_update.arg = READ_ONCE(sqe->addr);
5771 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
5773 struct io_ring_ctx *ctx = req->ctx;
5774 struct io_uring_rsrc_update up;
5777 if (issue_flags & IO_URING_F_NONBLOCK)
5780 up.offset = req->rsrc_update.offset;
5781 up.data = req->rsrc_update.arg;
5783 mutex_lock(&ctx->uring_lock);
5784 ret = __io_sqe_files_update(ctx, &up, req->rsrc_update.nr_args);
5785 mutex_unlock(&ctx->uring_lock);
5788 req_set_fail_links(req);
5789 __io_req_complete(req, issue_flags, ret, 0);
5793 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5795 switch (req->opcode) {
5798 case IORING_OP_READV:
5799 case IORING_OP_READ_FIXED:
5800 case IORING_OP_READ:
5801 return io_read_prep(req, sqe);
5802 case IORING_OP_WRITEV:
5803 case IORING_OP_WRITE_FIXED:
5804 case IORING_OP_WRITE:
5805 return io_write_prep(req, sqe);
5806 case IORING_OP_POLL_ADD:
5807 return io_poll_add_prep(req, sqe);
5808 case IORING_OP_POLL_REMOVE:
5809 return io_poll_remove_prep(req, sqe);
5810 case IORING_OP_FSYNC:
5811 return io_fsync_prep(req, sqe);
5812 case IORING_OP_SYNC_FILE_RANGE:
5813 return io_sfr_prep(req, sqe);
5814 case IORING_OP_SENDMSG:
5815 case IORING_OP_SEND:
5816 return io_sendmsg_prep(req, sqe);
5817 case IORING_OP_RECVMSG:
5818 case IORING_OP_RECV:
5819 return io_recvmsg_prep(req, sqe);
5820 case IORING_OP_CONNECT:
5821 return io_connect_prep(req, sqe);
5822 case IORING_OP_TIMEOUT:
5823 return io_timeout_prep(req, sqe, false);
5824 case IORING_OP_TIMEOUT_REMOVE:
5825 return io_timeout_remove_prep(req, sqe);
5826 case IORING_OP_ASYNC_CANCEL:
5827 return io_async_cancel_prep(req, sqe);
5828 case IORING_OP_LINK_TIMEOUT:
5829 return io_timeout_prep(req, sqe, true);
5830 case IORING_OP_ACCEPT:
5831 return io_accept_prep(req, sqe);
5832 case IORING_OP_FALLOCATE:
5833 return io_fallocate_prep(req, sqe);
5834 case IORING_OP_OPENAT:
5835 return io_openat_prep(req, sqe);
5836 case IORING_OP_CLOSE:
5837 return io_close_prep(req, sqe);
5838 case IORING_OP_FILES_UPDATE:
5839 return io_rsrc_update_prep(req, sqe);
5840 case IORING_OP_STATX:
5841 return io_statx_prep(req, sqe);
5842 case IORING_OP_FADVISE:
5843 return io_fadvise_prep(req, sqe);
5844 case IORING_OP_MADVISE:
5845 return io_madvise_prep(req, sqe);
5846 case IORING_OP_OPENAT2:
5847 return io_openat2_prep(req, sqe);
5848 case IORING_OP_EPOLL_CTL:
5849 return io_epoll_ctl_prep(req, sqe);
5850 case IORING_OP_SPLICE:
5851 return io_splice_prep(req, sqe);
5852 case IORING_OP_PROVIDE_BUFFERS:
5853 return io_provide_buffers_prep(req, sqe);
5854 case IORING_OP_REMOVE_BUFFERS:
5855 return io_remove_buffers_prep(req, sqe);
5857 return io_tee_prep(req, sqe);
5858 case IORING_OP_SHUTDOWN:
5859 return io_shutdown_prep(req, sqe);
5860 case IORING_OP_RENAMEAT:
5861 return io_renameat_prep(req, sqe);
5862 case IORING_OP_UNLINKAT:
5863 return io_unlinkat_prep(req, sqe);
5866 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5871 static int io_req_prep_async(struct io_kiocb *req)
5873 switch (req->opcode) {
5874 case IORING_OP_READV:
5875 case IORING_OP_READ_FIXED:
5876 case IORING_OP_READ:
5877 return io_rw_prep_async(req, READ);
5878 case IORING_OP_WRITEV:
5879 case IORING_OP_WRITE_FIXED:
5880 case IORING_OP_WRITE:
5881 return io_rw_prep_async(req, WRITE);
5882 case IORING_OP_SENDMSG:
5883 case IORING_OP_SEND:
5884 return io_sendmsg_prep_async(req);
5885 case IORING_OP_RECVMSG:
5886 case IORING_OP_RECV:
5887 return io_recvmsg_prep_async(req);
5888 case IORING_OP_CONNECT:
5889 return io_connect_prep_async(req);
5894 static int io_req_defer_prep(struct io_kiocb *req)
5896 if (!io_op_defs[req->opcode].needs_async_data)
5898 /* some opcodes init it during the inital prep */
5899 if (req->async_data)
5901 if (__io_alloc_async_data(req))
5903 return io_req_prep_async(req);
5906 static u32 io_get_sequence(struct io_kiocb *req)
5908 struct io_kiocb *pos;
5909 struct io_ring_ctx *ctx = req->ctx;
5910 u32 total_submitted, nr_reqs = 0;
5912 io_for_each_link(pos, req)
5915 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5916 return total_submitted - nr_reqs;
5919 static int io_req_defer(struct io_kiocb *req)
5921 struct io_ring_ctx *ctx = req->ctx;
5922 struct io_defer_entry *de;
5926 /* Still need defer if there is pending req in defer list. */
5927 if (likely(list_empty_careful(&ctx->defer_list) &&
5928 !(req->flags & REQ_F_IO_DRAIN)))
5931 seq = io_get_sequence(req);
5932 /* Still a chance to pass the sequence check */
5933 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
5936 ret = io_req_defer_prep(req);
5939 io_prep_async_link(req);
5940 de = kmalloc(sizeof(*de), GFP_KERNEL);
5944 spin_lock_irq(&ctx->completion_lock);
5945 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
5946 spin_unlock_irq(&ctx->completion_lock);
5948 io_queue_async_work(req);
5949 return -EIOCBQUEUED;
5952 trace_io_uring_defer(ctx, req, req->user_data);
5955 list_add_tail(&de->list, &ctx->defer_list);
5956 spin_unlock_irq(&ctx->completion_lock);
5957 return -EIOCBQUEUED;
5960 static void __io_clean_op(struct io_kiocb *req)
5962 if (req->flags & REQ_F_BUFFER_SELECTED) {
5963 switch (req->opcode) {
5964 case IORING_OP_READV:
5965 case IORING_OP_READ_FIXED:
5966 case IORING_OP_READ:
5967 kfree((void *)(unsigned long)req->rw.addr);
5969 case IORING_OP_RECVMSG:
5970 case IORING_OP_RECV:
5971 kfree(req->sr_msg.kbuf);
5974 req->flags &= ~REQ_F_BUFFER_SELECTED;
5977 if (req->flags & REQ_F_NEED_CLEANUP) {
5978 switch (req->opcode) {
5979 case IORING_OP_READV:
5980 case IORING_OP_READ_FIXED:
5981 case IORING_OP_READ:
5982 case IORING_OP_WRITEV:
5983 case IORING_OP_WRITE_FIXED:
5984 case IORING_OP_WRITE: {
5985 struct io_async_rw *io = req->async_data;
5987 kfree(io->free_iovec);
5990 case IORING_OP_RECVMSG:
5991 case IORING_OP_SENDMSG: {
5992 struct io_async_msghdr *io = req->async_data;
5994 kfree(io->free_iov);
5997 case IORING_OP_SPLICE:
5999 io_put_file(req, req->splice.file_in,
6000 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
6002 case IORING_OP_OPENAT:
6003 case IORING_OP_OPENAT2:
6004 if (req->open.filename)
6005 putname(req->open.filename);
6007 case IORING_OP_RENAMEAT:
6008 putname(req->rename.oldpath);
6009 putname(req->rename.newpath);
6011 case IORING_OP_UNLINKAT:
6012 putname(req->unlink.filename);
6015 req->flags &= ~REQ_F_NEED_CLEANUP;
6019 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
6021 struct io_ring_ctx *ctx = req->ctx;
6022 const struct cred *creds = NULL;
6025 if (req->work.creds && req->work.creds != current_cred())
6026 creds = override_creds(req->work.creds);
6028 switch (req->opcode) {
6030 ret = io_nop(req, issue_flags);
6032 case IORING_OP_READV:
6033 case IORING_OP_READ_FIXED:
6034 case IORING_OP_READ:
6035 ret = io_read(req, issue_flags);
6037 case IORING_OP_WRITEV:
6038 case IORING_OP_WRITE_FIXED:
6039 case IORING_OP_WRITE:
6040 ret = io_write(req, issue_flags);
6042 case IORING_OP_FSYNC:
6043 ret = io_fsync(req, issue_flags);
6045 case IORING_OP_POLL_ADD:
6046 ret = io_poll_add(req, issue_flags);
6048 case IORING_OP_POLL_REMOVE:
6049 ret = io_poll_remove(req, issue_flags);
6051 case IORING_OP_SYNC_FILE_RANGE:
6052 ret = io_sync_file_range(req, issue_flags);
6054 case IORING_OP_SENDMSG:
6055 ret = io_sendmsg(req, issue_flags);
6057 case IORING_OP_SEND:
6058 ret = io_send(req, issue_flags);
6060 case IORING_OP_RECVMSG:
6061 ret = io_recvmsg(req, issue_flags);
6063 case IORING_OP_RECV:
6064 ret = io_recv(req, issue_flags);
6066 case IORING_OP_TIMEOUT:
6067 ret = io_timeout(req, issue_flags);
6069 case IORING_OP_TIMEOUT_REMOVE:
6070 ret = io_timeout_remove(req, issue_flags);
6072 case IORING_OP_ACCEPT:
6073 ret = io_accept(req, issue_flags);
6075 case IORING_OP_CONNECT:
6076 ret = io_connect(req, issue_flags);
6078 case IORING_OP_ASYNC_CANCEL:
6079 ret = io_async_cancel(req, issue_flags);
6081 case IORING_OP_FALLOCATE:
6082 ret = io_fallocate(req, issue_flags);
6084 case IORING_OP_OPENAT:
6085 ret = io_openat(req, issue_flags);
6087 case IORING_OP_CLOSE:
6088 ret = io_close(req, issue_flags);
6090 case IORING_OP_FILES_UPDATE:
6091 ret = io_files_update(req, issue_flags);
6093 case IORING_OP_STATX:
6094 ret = io_statx(req, issue_flags);
6096 case IORING_OP_FADVISE:
6097 ret = io_fadvise(req, issue_flags);
6099 case IORING_OP_MADVISE:
6100 ret = io_madvise(req, issue_flags);
6102 case IORING_OP_OPENAT2:
6103 ret = io_openat2(req, issue_flags);
6105 case IORING_OP_EPOLL_CTL:
6106 ret = io_epoll_ctl(req, issue_flags);
6108 case IORING_OP_SPLICE:
6109 ret = io_splice(req, issue_flags);
6111 case IORING_OP_PROVIDE_BUFFERS:
6112 ret = io_provide_buffers(req, issue_flags);
6114 case IORING_OP_REMOVE_BUFFERS:
6115 ret = io_remove_buffers(req, issue_flags);
6118 ret = io_tee(req, issue_flags);
6120 case IORING_OP_SHUTDOWN:
6121 ret = io_shutdown(req, issue_flags);
6123 case IORING_OP_RENAMEAT:
6124 ret = io_renameat(req, issue_flags);
6126 case IORING_OP_UNLINKAT:
6127 ret = io_unlinkat(req, issue_flags);
6135 revert_creds(creds);
6140 /* If the op doesn't have a file, we're not polling for it */
6141 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
6142 const bool in_async = io_wq_current_is_worker();
6144 /* workqueue context doesn't hold uring_lock, grab it now */
6146 mutex_lock(&ctx->uring_lock);
6148 io_iopoll_req_issued(req, in_async);
6151 mutex_unlock(&ctx->uring_lock);
6157 static void io_wq_submit_work(struct io_wq_work *work)
6159 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6160 struct io_kiocb *timeout;
6163 timeout = io_prep_linked_timeout(req);
6165 io_queue_linked_timeout(timeout);
6167 if (work->flags & IO_WQ_WORK_CANCEL)
6171 req->flags &= ~REQ_F_REISSUE;
6173 ret = io_issue_sqe(req, 0);
6175 * We can get EAGAIN for polled IO even though we're
6176 * forcing a sync submission from here, since we can't
6177 * wait for request slots on the block side.
6185 /* avoid locking problems by failing it from a clean context */
6187 /* io-wq is going to take one down */
6188 refcount_inc(&req->refs);
6189 io_req_task_queue_fail(req, ret);
6193 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6196 struct fixed_rsrc_table *table;
6198 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
6199 return table->files[index & IORING_FILE_TABLE_MASK];
6202 static struct file *io_file_get(struct io_submit_state *state,
6203 struct io_kiocb *req, int fd, bool fixed)
6205 struct io_ring_ctx *ctx = req->ctx;
6209 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6211 fd = array_index_nospec(fd, ctx->nr_user_files);
6212 file = io_file_from_index(ctx, fd);
6213 io_set_resource_node(req);
6215 trace_io_uring_file_get(ctx, fd);
6216 file = __io_file_get(state, fd);
6219 if (file && unlikely(file->f_op == &io_uring_fops))
6220 io_req_track_inflight(req);
6224 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6226 struct io_timeout_data *data = container_of(timer,
6227 struct io_timeout_data, timer);
6228 struct io_kiocb *prev, *req = data->req;
6229 struct io_ring_ctx *ctx = req->ctx;
6230 unsigned long flags;
6232 spin_lock_irqsave(&ctx->completion_lock, flags);
6233 prev = req->timeout.head;
6234 req->timeout.head = NULL;
6237 * We don't expect the list to be empty, that will only happen if we
6238 * race with the completion of the linked work.
6240 if (prev && refcount_inc_not_zero(&prev->refs))
6241 io_remove_next_linked(prev);
6244 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6247 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6248 io_put_req_deferred(prev, 1);
6250 io_req_complete_post(req, -ETIME, 0);
6251 io_put_req_deferred(req, 1);
6253 return HRTIMER_NORESTART;
6256 static void __io_queue_linked_timeout(struct io_kiocb *req)
6259 * If the back reference is NULL, then our linked request finished
6260 * before we got a chance to setup the timer
6262 if (req->timeout.head) {
6263 struct io_timeout_data *data = req->async_data;
6265 data->timer.function = io_link_timeout_fn;
6266 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6271 static void io_queue_linked_timeout(struct io_kiocb *req)
6273 struct io_ring_ctx *ctx = req->ctx;
6275 spin_lock_irq(&ctx->completion_lock);
6276 __io_queue_linked_timeout(req);
6277 spin_unlock_irq(&ctx->completion_lock);
6279 /* drop submission reference */
6283 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6285 struct io_kiocb *nxt = req->link;
6287 if (!nxt || (req->flags & REQ_F_LINK_TIMEOUT) ||
6288 nxt->opcode != IORING_OP_LINK_TIMEOUT)
6291 nxt->timeout.head = req;
6292 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6293 req->flags |= REQ_F_LINK_TIMEOUT;
6297 static void __io_queue_sqe(struct io_kiocb *req)
6299 struct io_kiocb *linked_timeout = io_prep_linked_timeout(req);
6302 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6305 * We async punt it if the file wasn't marked NOWAIT, or if the file
6306 * doesn't support non-blocking read/write attempts
6308 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6309 if (!io_arm_poll_handler(req)) {
6311 * Queued up for async execution, worker will release
6312 * submit reference when the iocb is actually submitted.
6314 io_queue_async_work(req);
6316 } else if (likely(!ret)) {
6317 /* drop submission reference */
6318 if (req->flags & REQ_F_COMPLETE_INLINE) {
6319 struct io_ring_ctx *ctx = req->ctx;
6320 struct io_comp_state *cs = &ctx->submit_state.comp;
6322 cs->reqs[cs->nr++] = req;
6323 if (cs->nr == ARRAY_SIZE(cs->reqs))
6324 io_submit_flush_completions(cs, ctx);
6329 req_set_fail_links(req);
6331 io_req_complete(req, ret);
6334 io_queue_linked_timeout(linked_timeout);
6337 static void io_queue_sqe(struct io_kiocb *req)
6341 ret = io_req_defer(req);
6343 if (ret != -EIOCBQUEUED) {
6345 req_set_fail_links(req);
6347 io_req_complete(req, ret);
6349 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6350 ret = io_req_defer_prep(req);
6353 io_queue_async_work(req);
6355 __io_queue_sqe(req);
6360 * Check SQE restrictions (opcode and flags).
6362 * Returns 'true' if SQE is allowed, 'false' otherwise.
6364 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6365 struct io_kiocb *req,
6366 unsigned int sqe_flags)
6368 if (!ctx->restricted)
6371 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6374 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6375 ctx->restrictions.sqe_flags_required)
6378 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6379 ctx->restrictions.sqe_flags_required))
6385 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6386 const struct io_uring_sqe *sqe)
6388 struct io_submit_state *state;
6389 unsigned int sqe_flags;
6390 int personality, ret = 0;
6392 req->opcode = READ_ONCE(sqe->opcode);
6393 /* same numerical values with corresponding REQ_F_*, safe to copy */
6394 req->flags = sqe_flags = READ_ONCE(sqe->flags);
6395 req->user_data = READ_ONCE(sqe->user_data);
6396 req->async_data = NULL;
6400 req->fixed_rsrc_refs = NULL;
6401 /* one is dropped after submission, the other at completion */
6402 refcount_set(&req->refs, 2);
6403 req->task = current;
6405 req->work.list.next = NULL;
6406 req->work.creds = NULL;
6407 req->work.flags = 0;
6409 /* enforce forwards compatibility on users */
6410 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS)) {
6415 if (unlikely(req->opcode >= IORING_OP_LAST))
6418 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6421 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6422 !io_op_defs[req->opcode].buffer_select)
6425 personality = READ_ONCE(sqe->personality);
6427 req->work.creds = xa_load(&ctx->personalities, personality);
6428 if (!req->work.creds)
6430 get_cred(req->work.creds);
6432 state = &ctx->submit_state;
6435 * Plug now if we have more than 1 IO left after this, and the target
6436 * is potentially a read/write to block based storage.
6438 if (!state->plug_started && state->ios_left > 1 &&
6439 io_op_defs[req->opcode].plug) {
6440 blk_start_plug(&state->plug);
6441 state->plug_started = true;
6444 if (io_op_defs[req->opcode].needs_file) {
6445 bool fixed = req->flags & REQ_F_FIXED_FILE;
6447 req->file = io_file_get(state, req, READ_ONCE(sqe->fd), fixed);
6448 if (unlikely(!req->file))
6456 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
6457 const struct io_uring_sqe *sqe)
6459 struct io_submit_link *link = &ctx->submit_state.link;
6462 ret = io_init_req(ctx, req, sqe);
6463 if (unlikely(ret)) {
6466 /* fail even hard links since we don't submit */
6467 link->head->flags |= REQ_F_FAIL_LINK;
6468 io_put_req(link->head);
6469 io_req_complete(link->head, -ECANCELED);
6473 io_req_complete(req, ret);
6476 ret = io_req_prep(req, sqe);
6480 /* don't need @sqe from now on */
6481 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6482 true, ctx->flags & IORING_SETUP_SQPOLL);
6485 * If we already have a head request, queue this one for async
6486 * submittal once the head completes. If we don't have a head but
6487 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6488 * submitted sync once the chain is complete. If none of those
6489 * conditions are true (normal request), then just queue it.
6492 struct io_kiocb *head = link->head;
6495 * Taking sequential execution of a link, draining both sides
6496 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6497 * requests in the link. So, it drains the head and the
6498 * next after the link request. The last one is done via
6499 * drain_next flag to persist the effect across calls.
6501 if (req->flags & REQ_F_IO_DRAIN) {
6502 head->flags |= REQ_F_IO_DRAIN;
6503 ctx->drain_next = 1;
6505 ret = io_req_defer_prep(req);
6508 trace_io_uring_link(ctx, req, head);
6509 link->last->link = req;
6512 /* last request of a link, enqueue the link */
6513 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6518 if (unlikely(ctx->drain_next)) {
6519 req->flags |= REQ_F_IO_DRAIN;
6520 ctx->drain_next = 0;
6522 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6534 * Batched submission is done, ensure local IO is flushed out.
6536 static void io_submit_state_end(struct io_submit_state *state,
6537 struct io_ring_ctx *ctx)
6539 if (state->link.head)
6540 io_queue_sqe(state->link.head);
6542 io_submit_flush_completions(&state->comp, ctx);
6543 if (state->plug_started)
6544 blk_finish_plug(&state->plug);
6545 io_state_file_put(state);
6549 * Start submission side cache.
6551 static void io_submit_state_start(struct io_submit_state *state,
6552 unsigned int max_ios)
6554 state->plug_started = false;
6555 state->ios_left = max_ios;
6556 /* set only head, no need to init link_last in advance */
6557 state->link.head = NULL;
6560 static void io_commit_sqring(struct io_ring_ctx *ctx)
6562 struct io_rings *rings = ctx->rings;
6565 * Ensure any loads from the SQEs are done at this point,
6566 * since once we write the new head, the application could
6567 * write new data to them.
6569 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6573 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6574 * that is mapped by userspace. This means that care needs to be taken to
6575 * ensure that reads are stable, as we cannot rely on userspace always
6576 * being a good citizen. If members of the sqe are validated and then later
6577 * used, it's important that those reads are done through READ_ONCE() to
6578 * prevent a re-load down the line.
6580 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6582 u32 *sq_array = ctx->sq_array;
6586 * The cached sq head (or cq tail) serves two purposes:
6588 * 1) allows us to batch the cost of updating the user visible
6590 * 2) allows the kernel side to track the head on its own, even
6591 * though the application is the one updating it.
6593 head = READ_ONCE(sq_array[ctx->cached_sq_head++ & ctx->sq_mask]);
6594 if (likely(head < ctx->sq_entries))
6595 return &ctx->sq_sqes[head];
6597 /* drop invalid entries */
6598 ctx->cached_sq_dropped++;
6599 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6603 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6607 /* if we have a backlog and couldn't flush it all, return BUSY */
6608 if (test_bit(0, &ctx->sq_check_overflow)) {
6609 if (!__io_cqring_overflow_flush(ctx, false, NULL, NULL))
6613 /* make sure SQ entry isn't read before tail */
6614 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6616 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6619 percpu_counter_add(¤t->io_uring->inflight, nr);
6620 refcount_add(nr, ¤t->usage);
6621 io_submit_state_start(&ctx->submit_state, nr);
6623 while (submitted < nr) {
6624 const struct io_uring_sqe *sqe;
6625 struct io_kiocb *req;
6627 req = io_alloc_req(ctx);
6628 if (unlikely(!req)) {
6630 submitted = -EAGAIN;
6633 sqe = io_get_sqe(ctx);
6634 if (unlikely(!sqe)) {
6635 kmem_cache_free(req_cachep, req);
6638 /* will complete beyond this point, count as submitted */
6640 if (io_submit_sqe(ctx, req, sqe))
6644 if (unlikely(submitted != nr)) {
6645 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6646 struct io_uring_task *tctx = current->io_uring;
6647 int unused = nr - ref_used;
6649 percpu_ref_put_many(&ctx->refs, unused);
6650 percpu_counter_sub(&tctx->inflight, unused);
6651 put_task_struct_many(current, unused);
6654 io_submit_state_end(&ctx->submit_state, ctx);
6655 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6656 io_commit_sqring(ctx);
6661 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6663 /* Tell userspace we may need a wakeup call */
6664 spin_lock_irq(&ctx->completion_lock);
6665 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6666 spin_unlock_irq(&ctx->completion_lock);
6669 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6671 spin_lock_irq(&ctx->completion_lock);
6672 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6673 spin_unlock_irq(&ctx->completion_lock);
6676 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6678 unsigned int to_submit;
6681 to_submit = io_sqring_entries(ctx);
6682 /* if we're handling multiple rings, cap submit size for fairness */
6683 if (cap_entries && to_submit > 8)
6686 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6687 unsigned nr_events = 0;
6689 mutex_lock(&ctx->uring_lock);
6690 if (!list_empty(&ctx->iopoll_list))
6691 io_do_iopoll(ctx, &nr_events, 0);
6693 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
6694 !(ctx->flags & IORING_SETUP_R_DISABLED))
6695 ret = io_submit_sqes(ctx, to_submit);
6696 mutex_unlock(&ctx->uring_lock);
6699 if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
6700 wake_up(&ctx->sqo_sq_wait);
6705 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
6707 struct io_ring_ctx *ctx;
6708 unsigned sq_thread_idle = 0;
6710 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6711 if (sq_thread_idle < ctx->sq_thread_idle)
6712 sq_thread_idle = ctx->sq_thread_idle;
6715 sqd->sq_thread_idle = sq_thread_idle;
6718 static int io_sq_thread(void *data)
6720 struct io_sq_data *sqd = data;
6721 struct io_ring_ctx *ctx;
6722 unsigned long timeout = 0;
6723 char buf[TASK_COMM_LEN];
6726 snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
6727 set_task_comm(current, buf);
6728 current->pf_io_worker = NULL;
6730 if (sqd->sq_cpu != -1)
6731 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
6733 set_cpus_allowed_ptr(current, cpu_online_mask);
6734 current->flags |= PF_NO_SETAFFINITY;
6736 mutex_lock(&sqd->lock);
6737 while (!test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state)) {
6739 bool cap_entries, sqt_spin, needs_sched;
6741 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) ||
6742 signal_pending(current)) {
6743 bool did_sig = false;
6745 mutex_unlock(&sqd->lock);
6746 if (signal_pending(current)) {
6747 struct ksignal ksig;
6749 did_sig = get_signal(&ksig);
6752 mutex_lock(&sqd->lock);
6756 io_run_task_work_head(&sqd->park_task_work);
6757 timeout = jiffies + sqd->sq_thread_idle;
6761 cap_entries = !list_is_singular(&sqd->ctx_list);
6762 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6763 const struct cred *creds = NULL;
6765 if (ctx->sq_creds != current_cred())
6766 creds = override_creds(ctx->sq_creds);
6767 ret = __io_sq_thread(ctx, cap_entries);
6769 revert_creds(creds);
6770 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
6774 if (sqt_spin || !time_after(jiffies, timeout)) {
6778 timeout = jiffies + sqd->sq_thread_idle;
6783 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
6784 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6785 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6786 !list_empty_careful(&ctx->iopoll_list)) {
6787 needs_sched = false;
6790 if (io_sqring_entries(ctx)) {
6791 needs_sched = false;
6796 if (needs_sched && !test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state)) {
6797 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6798 io_ring_set_wakeup_flag(ctx);
6800 mutex_unlock(&sqd->lock);
6802 mutex_lock(&sqd->lock);
6803 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6804 io_ring_clear_wakeup_flag(ctx);
6807 finish_wait(&sqd->wait, &wait);
6808 io_run_task_work_head(&sqd->park_task_work);
6809 timeout = jiffies + sqd->sq_thread_idle;
6812 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6813 io_uring_cancel_sqpoll(ctx);
6815 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6816 io_ring_set_wakeup_flag(ctx);
6817 mutex_unlock(&sqd->lock);
6820 io_run_task_work_head(&sqd->park_task_work);
6821 complete(&sqd->exited);
6825 struct io_wait_queue {
6826 struct wait_queue_entry wq;
6827 struct io_ring_ctx *ctx;
6829 unsigned nr_timeouts;
6832 static inline bool io_should_wake(struct io_wait_queue *iowq)
6834 struct io_ring_ctx *ctx = iowq->ctx;
6837 * Wake up if we have enough events, or if a timeout occurred since we
6838 * started waiting. For timeouts, we always want to return to userspace,
6839 * regardless of event count.
6841 return io_cqring_events(ctx) >= iowq->to_wait ||
6842 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6845 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6846 int wake_flags, void *key)
6848 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6852 * Cannot safely flush overflowed CQEs from here, ensure we wake up
6853 * the task, and the next invocation will do it.
6855 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->cq_check_overflow))
6856 return autoremove_wake_function(curr, mode, wake_flags, key);
6860 static int io_run_task_work_sig(void)
6862 if (io_run_task_work())
6864 if (!signal_pending(current))
6866 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
6867 return -ERESTARTSYS;
6871 /* when returns >0, the caller should retry */
6872 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
6873 struct io_wait_queue *iowq,
6874 signed long *timeout)
6878 /* make sure we run task_work before checking for signals */
6879 ret = io_run_task_work_sig();
6880 if (ret || io_should_wake(iowq))
6882 /* let the caller flush overflows, retry */
6883 if (test_bit(0, &ctx->cq_check_overflow))
6886 *timeout = schedule_timeout(*timeout);
6887 return !*timeout ? -ETIME : 1;
6891 * Wait until events become available, if we don't already have some. The
6892 * application must reap them itself, as they reside on the shared cq ring.
6894 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6895 const sigset_t __user *sig, size_t sigsz,
6896 struct __kernel_timespec __user *uts)
6898 struct io_wait_queue iowq = {
6901 .func = io_wake_function,
6902 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6905 .to_wait = min_events,
6907 struct io_rings *rings = ctx->rings;
6908 signed long timeout = MAX_SCHEDULE_TIMEOUT;
6912 io_cqring_overflow_flush(ctx, false, NULL, NULL);
6913 if (io_cqring_events(ctx) >= min_events)
6915 if (!io_run_task_work())
6920 #ifdef CONFIG_COMPAT
6921 if (in_compat_syscall())
6922 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6926 ret = set_user_sigmask(sig, sigsz);
6933 struct timespec64 ts;
6935 if (get_timespec64(&ts, uts))
6937 timeout = timespec64_to_jiffies(&ts);
6940 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6941 trace_io_uring_cqring_wait(ctx, min_events);
6943 /* if we can't even flush overflow, don't wait for more */
6944 if (!io_cqring_overflow_flush(ctx, false, NULL, NULL)) {
6948 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6949 TASK_INTERRUPTIBLE);
6950 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
6951 finish_wait(&ctx->wait, &iowq.wq);
6955 restore_saved_sigmask_unless(ret == -EINTR);
6957 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6960 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6962 #if defined(CONFIG_UNIX)
6963 if (ctx->ring_sock) {
6964 struct sock *sock = ctx->ring_sock->sk;
6965 struct sk_buff *skb;
6967 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6973 for (i = 0; i < ctx->nr_user_files; i++) {
6976 file = io_file_from_index(ctx, i);
6983 static void io_rsrc_data_ref_zero(struct percpu_ref *ref)
6985 struct fixed_rsrc_data *data;
6987 data = container_of(ref, struct fixed_rsrc_data, refs);
6988 complete(&data->done);
6991 static inline void io_rsrc_ref_lock(struct io_ring_ctx *ctx)
6993 spin_lock_bh(&ctx->rsrc_ref_lock);
6996 static inline void io_rsrc_ref_unlock(struct io_ring_ctx *ctx)
6998 spin_unlock_bh(&ctx->rsrc_ref_lock);
7001 static void io_sqe_rsrc_set_node(struct io_ring_ctx *ctx,
7002 struct fixed_rsrc_data *rsrc_data,
7003 struct fixed_rsrc_ref_node *ref_node)
7005 io_rsrc_ref_lock(ctx);
7006 rsrc_data->node = ref_node;
7007 list_add_tail(&ref_node->node, &ctx->rsrc_ref_list);
7008 io_rsrc_ref_unlock(ctx);
7009 percpu_ref_get(&rsrc_data->refs);
7012 static void io_sqe_rsrc_kill_node(struct io_ring_ctx *ctx, struct fixed_rsrc_data *data)
7014 struct fixed_rsrc_ref_node *ref_node = NULL;
7016 io_rsrc_ref_lock(ctx);
7017 ref_node = data->node;
7019 io_rsrc_ref_unlock(ctx);
7021 percpu_ref_kill(&ref_node->refs);
7024 static int io_rsrc_ref_quiesce(struct fixed_rsrc_data *data,
7025 struct io_ring_ctx *ctx,
7026 void (*rsrc_put)(struct io_ring_ctx *ctx,
7027 struct io_rsrc_put *prsrc))
7029 struct fixed_rsrc_ref_node *backup_node;
7035 data->quiesce = true;
7038 backup_node = alloc_fixed_rsrc_ref_node(ctx);
7041 backup_node->rsrc_data = data;
7042 backup_node->rsrc_put = rsrc_put;
7044 io_sqe_rsrc_kill_node(ctx, data);
7045 percpu_ref_kill(&data->refs);
7046 flush_delayed_work(&ctx->rsrc_put_work);
7048 ret = wait_for_completion_interruptible(&data->done);
7052 percpu_ref_resurrect(&data->refs);
7053 io_sqe_rsrc_set_node(ctx, data, backup_node);
7055 reinit_completion(&data->done);
7056 mutex_unlock(&ctx->uring_lock);
7057 ret = io_run_task_work_sig();
7058 mutex_lock(&ctx->uring_lock);
7060 data->quiesce = false;
7063 destroy_fixed_rsrc_ref_node(backup_node);
7067 static struct fixed_rsrc_data *alloc_fixed_rsrc_data(struct io_ring_ctx *ctx)
7069 struct fixed_rsrc_data *data;
7071 data = kzalloc(sizeof(*data), GFP_KERNEL);
7075 if (percpu_ref_init(&data->refs, io_rsrc_data_ref_zero,
7076 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
7081 init_completion(&data->done);
7085 static void free_fixed_rsrc_data(struct fixed_rsrc_data *data)
7087 percpu_ref_exit(&data->refs);
7092 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7094 struct fixed_rsrc_data *data = ctx->file_data;
7095 unsigned nr_tables, i;
7099 * percpu_ref_is_dying() is to stop parallel files unregister
7100 * Since we possibly drop uring lock later in this function to
7103 if (!data || percpu_ref_is_dying(&data->refs))
7105 ret = io_rsrc_ref_quiesce(data, ctx, io_ring_file_put);
7109 __io_sqe_files_unregister(ctx);
7110 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
7111 for (i = 0; i < nr_tables; i++)
7112 kfree(data->table[i].files);
7113 free_fixed_rsrc_data(data);
7114 ctx->file_data = NULL;
7115 ctx->nr_user_files = 0;
7119 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7120 __releases(&sqd->lock)
7122 WARN_ON_ONCE(sqd->thread == current);
7125 * Do the dance but not conditional clear_bit() because it'd race with
7126 * other threads incrementing park_pending and setting the bit.
7128 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7129 if (atomic_dec_return(&sqd->park_pending))
7130 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7131 mutex_unlock(&sqd->lock);
7134 static void io_sq_thread_park(struct io_sq_data *sqd)
7135 __acquires(&sqd->lock)
7137 WARN_ON_ONCE(sqd->thread == current);
7139 atomic_inc(&sqd->park_pending);
7140 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7141 mutex_lock(&sqd->lock);
7143 wake_up_process(sqd->thread);
7146 static void io_sq_thread_stop(struct io_sq_data *sqd)
7148 WARN_ON_ONCE(sqd->thread == current);
7150 mutex_lock(&sqd->lock);
7151 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7153 wake_up_process(sqd->thread);
7154 mutex_unlock(&sqd->lock);
7155 wait_for_completion(&sqd->exited);
7158 static void io_put_sq_data(struct io_sq_data *sqd)
7160 if (refcount_dec_and_test(&sqd->refs)) {
7161 WARN_ON_ONCE(atomic_read(&sqd->park_pending));
7163 io_sq_thread_stop(sqd);
7168 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7170 struct io_sq_data *sqd = ctx->sq_data;
7173 io_sq_thread_park(sqd);
7174 list_del_init(&ctx->sqd_list);
7175 io_sqd_update_thread_idle(sqd);
7176 io_sq_thread_unpark(sqd);
7178 io_put_sq_data(sqd);
7179 ctx->sq_data = NULL;
7181 put_cred(ctx->sq_creds);
7185 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7187 struct io_ring_ctx *ctx_attach;
7188 struct io_sq_data *sqd;
7191 f = fdget(p->wq_fd);
7193 return ERR_PTR(-ENXIO);
7194 if (f.file->f_op != &io_uring_fops) {
7196 return ERR_PTR(-EINVAL);
7199 ctx_attach = f.file->private_data;
7200 sqd = ctx_attach->sq_data;
7203 return ERR_PTR(-EINVAL);
7205 if (sqd->task_tgid != current->tgid) {
7207 return ERR_PTR(-EPERM);
7210 refcount_inc(&sqd->refs);
7215 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
7218 struct io_sq_data *sqd;
7221 if (p->flags & IORING_SETUP_ATTACH_WQ) {
7222 sqd = io_attach_sq_data(p);
7227 /* fall through for EPERM case, setup new sqd/task */
7228 if (PTR_ERR(sqd) != -EPERM)
7232 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7234 return ERR_PTR(-ENOMEM);
7236 atomic_set(&sqd->park_pending, 0);
7237 refcount_set(&sqd->refs, 1);
7238 INIT_LIST_HEAD(&sqd->ctx_list);
7239 mutex_init(&sqd->lock);
7240 init_waitqueue_head(&sqd->wait);
7241 init_completion(&sqd->exited);
7245 #if defined(CONFIG_UNIX)
7247 * Ensure the UNIX gc is aware of our file set, so we are certain that
7248 * the io_uring can be safely unregistered on process exit, even if we have
7249 * loops in the file referencing.
7251 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7253 struct sock *sk = ctx->ring_sock->sk;
7254 struct scm_fp_list *fpl;
7255 struct sk_buff *skb;
7258 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7262 skb = alloc_skb(0, GFP_KERNEL);
7271 fpl->user = get_uid(current_user());
7272 for (i = 0; i < nr; i++) {
7273 struct file *file = io_file_from_index(ctx, i + offset);
7277 fpl->fp[nr_files] = get_file(file);
7278 unix_inflight(fpl->user, fpl->fp[nr_files]);
7283 fpl->max = SCM_MAX_FD;
7284 fpl->count = nr_files;
7285 UNIXCB(skb).fp = fpl;
7286 skb->destructor = unix_destruct_scm;
7287 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7288 skb_queue_head(&sk->sk_receive_queue, skb);
7290 for (i = 0; i < nr_files; i++)
7301 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7302 * causes regular reference counting to break down. We rely on the UNIX
7303 * garbage collection to take care of this problem for us.
7305 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7307 unsigned left, total;
7311 left = ctx->nr_user_files;
7313 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7315 ret = __io_sqe_files_scm(ctx, this_files, total);
7319 total += this_files;
7325 while (total < ctx->nr_user_files) {
7326 struct file *file = io_file_from_index(ctx, total);
7336 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7342 static int io_sqe_alloc_file_tables(struct fixed_rsrc_data *file_data,
7343 unsigned nr_tables, unsigned nr_files)
7347 for (i = 0; i < nr_tables; i++) {
7348 struct fixed_rsrc_table *table = &file_data->table[i];
7349 unsigned this_files;
7351 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7352 table->files = kcalloc(this_files, sizeof(struct file *),
7356 nr_files -= this_files;
7362 for (i = 0; i < nr_tables; i++) {
7363 struct fixed_rsrc_table *table = &file_data->table[i];
7364 kfree(table->files);
7369 static void io_ring_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
7371 struct file *file = prsrc->file;
7372 #if defined(CONFIG_UNIX)
7373 struct sock *sock = ctx->ring_sock->sk;
7374 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7375 struct sk_buff *skb;
7378 __skb_queue_head_init(&list);
7381 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7382 * remove this entry and rearrange the file array.
7384 skb = skb_dequeue(head);
7386 struct scm_fp_list *fp;
7388 fp = UNIXCB(skb).fp;
7389 for (i = 0; i < fp->count; i++) {
7392 if (fp->fp[i] != file)
7395 unix_notinflight(fp->user, fp->fp[i]);
7396 left = fp->count - 1 - i;
7398 memmove(&fp->fp[i], &fp->fp[i + 1],
7399 left * sizeof(struct file *));
7406 __skb_queue_tail(&list, skb);
7416 __skb_queue_tail(&list, skb);
7418 skb = skb_dequeue(head);
7421 if (skb_peek(&list)) {
7422 spin_lock_irq(&head->lock);
7423 while ((skb = __skb_dequeue(&list)) != NULL)
7424 __skb_queue_tail(head, skb);
7425 spin_unlock_irq(&head->lock);
7432 static void __io_rsrc_put_work(struct fixed_rsrc_ref_node *ref_node)
7434 struct fixed_rsrc_data *rsrc_data = ref_node->rsrc_data;
7435 struct io_ring_ctx *ctx = rsrc_data->ctx;
7436 struct io_rsrc_put *prsrc, *tmp;
7438 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
7439 list_del(&prsrc->list);
7440 ref_node->rsrc_put(ctx, prsrc);
7444 percpu_ref_exit(&ref_node->refs);
7446 percpu_ref_put(&rsrc_data->refs);
7449 static void io_rsrc_put_work(struct work_struct *work)
7451 struct io_ring_ctx *ctx;
7452 struct llist_node *node;
7454 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
7455 node = llist_del_all(&ctx->rsrc_put_llist);
7458 struct fixed_rsrc_ref_node *ref_node;
7459 struct llist_node *next = node->next;
7461 ref_node = llist_entry(node, struct fixed_rsrc_ref_node, llist);
7462 __io_rsrc_put_work(ref_node);
7467 static struct file **io_fixed_file_slot(struct fixed_rsrc_data *file_data,
7470 struct fixed_rsrc_table *table;
7472 table = &file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7473 return &table->files[i & IORING_FILE_TABLE_MASK];
7476 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7478 struct fixed_rsrc_ref_node *ref_node;
7479 struct fixed_rsrc_data *data;
7480 struct io_ring_ctx *ctx;
7481 bool first_add = false;
7484 ref_node = container_of(ref, struct fixed_rsrc_ref_node, refs);
7485 data = ref_node->rsrc_data;
7488 io_rsrc_ref_lock(ctx);
7489 ref_node->done = true;
7491 while (!list_empty(&ctx->rsrc_ref_list)) {
7492 ref_node = list_first_entry(&ctx->rsrc_ref_list,
7493 struct fixed_rsrc_ref_node, node);
7494 /* recycle ref nodes in order */
7495 if (!ref_node->done)
7497 list_del(&ref_node->node);
7498 first_add |= llist_add(&ref_node->llist, &ctx->rsrc_put_llist);
7500 io_rsrc_ref_unlock(ctx);
7502 if (percpu_ref_is_dying(&data->refs))
7506 mod_delayed_work(system_wq, &ctx->rsrc_put_work, 0);
7508 queue_delayed_work(system_wq, &ctx->rsrc_put_work, delay);
7511 static struct fixed_rsrc_ref_node *alloc_fixed_rsrc_ref_node(
7512 struct io_ring_ctx *ctx)
7514 struct fixed_rsrc_ref_node *ref_node;
7516 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7520 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7525 INIT_LIST_HEAD(&ref_node->node);
7526 INIT_LIST_HEAD(&ref_node->rsrc_list);
7527 ref_node->done = false;
7531 static void init_fixed_file_ref_node(struct io_ring_ctx *ctx,
7532 struct fixed_rsrc_ref_node *ref_node)
7534 ref_node->rsrc_data = ctx->file_data;
7535 ref_node->rsrc_put = io_ring_file_put;
7538 static void destroy_fixed_rsrc_ref_node(struct fixed_rsrc_ref_node *ref_node)
7540 percpu_ref_exit(&ref_node->refs);
7545 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7548 __s32 __user *fds = (__s32 __user *) arg;
7549 unsigned nr_tables, i;
7551 int fd, ret = -ENOMEM;
7552 struct fixed_rsrc_ref_node *ref_node;
7553 struct fixed_rsrc_data *file_data;
7559 if (nr_args > IORING_MAX_FIXED_FILES)
7562 file_data = alloc_fixed_rsrc_data(ctx);
7565 ctx->file_data = file_data;
7567 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7568 file_data->table = kcalloc(nr_tables, sizeof(*file_data->table),
7570 if (!file_data->table)
7573 if (io_sqe_alloc_file_tables(file_data, nr_tables, nr_args))
7576 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7577 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7581 /* allow sparse sets */
7591 * Don't allow io_uring instances to be registered. If UNIX
7592 * isn't enabled, then this causes a reference cycle and this
7593 * instance can never get freed. If UNIX is enabled we'll
7594 * handle it just fine, but there's still no point in allowing
7595 * a ring fd as it doesn't support regular read/write anyway.
7597 if (file->f_op == &io_uring_fops) {
7601 *io_fixed_file_slot(file_data, i) = file;
7604 ret = io_sqe_files_scm(ctx);
7606 io_sqe_files_unregister(ctx);
7610 ref_node = alloc_fixed_rsrc_ref_node(ctx);
7612 io_sqe_files_unregister(ctx);
7615 init_fixed_file_ref_node(ctx, ref_node);
7617 io_sqe_rsrc_set_node(ctx, file_data, ref_node);
7620 for (i = 0; i < ctx->nr_user_files; i++) {
7621 file = io_file_from_index(ctx, i);
7625 for (i = 0; i < nr_tables; i++)
7626 kfree(file_data->table[i].files);
7627 ctx->nr_user_files = 0;
7629 free_fixed_rsrc_data(ctx->file_data);
7630 ctx->file_data = NULL;
7634 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7637 #if defined(CONFIG_UNIX)
7638 struct sock *sock = ctx->ring_sock->sk;
7639 struct sk_buff_head *head = &sock->sk_receive_queue;
7640 struct sk_buff *skb;
7643 * See if we can merge this file into an existing skb SCM_RIGHTS
7644 * file set. If there's no room, fall back to allocating a new skb
7645 * and filling it in.
7647 spin_lock_irq(&head->lock);
7648 skb = skb_peek(head);
7650 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7652 if (fpl->count < SCM_MAX_FD) {
7653 __skb_unlink(skb, head);
7654 spin_unlock_irq(&head->lock);
7655 fpl->fp[fpl->count] = get_file(file);
7656 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7658 spin_lock_irq(&head->lock);
7659 __skb_queue_head(head, skb);
7664 spin_unlock_irq(&head->lock);
7671 return __io_sqe_files_scm(ctx, 1, index);
7677 static int io_queue_rsrc_removal(struct fixed_rsrc_data *data, void *rsrc)
7679 struct io_rsrc_put *prsrc;
7680 struct fixed_rsrc_ref_node *ref_node = data->node;
7682 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
7687 list_add(&prsrc->list, &ref_node->rsrc_list);
7692 static inline int io_queue_file_removal(struct fixed_rsrc_data *data,
7695 return io_queue_rsrc_removal(data, (void *)file);
7698 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7699 struct io_uring_rsrc_update *up,
7702 struct fixed_rsrc_data *data = ctx->file_data;
7703 struct fixed_rsrc_ref_node *ref_node;
7704 struct file *file, **file_slot;
7708 bool needs_switch = false;
7710 if (check_add_overflow(up->offset, nr_args, &done))
7712 if (done > ctx->nr_user_files)
7715 ref_node = alloc_fixed_rsrc_ref_node(ctx);
7718 init_fixed_file_ref_node(ctx, ref_node);
7720 fds = u64_to_user_ptr(up->data);
7721 for (done = 0; done < nr_args; done++) {
7723 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7727 if (fd == IORING_REGISTER_FILES_SKIP)
7730 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
7731 file_slot = io_fixed_file_slot(ctx->file_data, i);
7734 err = io_queue_file_removal(data, *file_slot);
7738 needs_switch = true;
7747 * Don't allow io_uring instances to be registered. If
7748 * UNIX isn't enabled, then this causes a reference
7749 * cycle and this instance can never get freed. If UNIX
7750 * is enabled we'll handle it just fine, but there's
7751 * still no point in allowing a ring fd as it doesn't
7752 * support regular read/write anyway.
7754 if (file->f_op == &io_uring_fops) {
7760 err = io_sqe_file_register(ctx, file, i);
7770 percpu_ref_kill(&data->node->refs);
7771 io_sqe_rsrc_set_node(ctx, data, ref_node);
7773 destroy_fixed_rsrc_ref_node(ref_node);
7775 return done ? done : err;
7778 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7781 struct io_uring_rsrc_update up;
7783 if (!ctx->file_data)
7787 if (copy_from_user(&up, arg, sizeof(up)))
7792 return __io_sqe_files_update(ctx, &up, nr_args);
7795 static struct io_wq_work *io_free_work(struct io_wq_work *work)
7797 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7799 req = io_put_req_find_next(req);
7800 return req ? &req->work : NULL;
7803 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx)
7805 struct io_wq_hash *hash;
7806 struct io_wq_data data;
7807 unsigned int concurrency;
7809 hash = ctx->hash_map;
7811 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
7813 return ERR_PTR(-ENOMEM);
7814 refcount_set(&hash->refs, 1);
7815 init_waitqueue_head(&hash->wait);
7816 ctx->hash_map = hash;
7820 data.free_work = io_free_work;
7821 data.do_work = io_wq_submit_work;
7823 /* Do QD, or 4 * CPUS, whatever is smallest */
7824 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7826 return io_wq_create(concurrency, &data);
7829 static int io_uring_alloc_task_context(struct task_struct *task,
7830 struct io_ring_ctx *ctx)
7832 struct io_uring_task *tctx;
7835 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
7836 if (unlikely(!tctx))
7839 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7840 if (unlikely(ret)) {
7845 tctx->io_wq = io_init_wq_offload(ctx);
7846 if (IS_ERR(tctx->io_wq)) {
7847 ret = PTR_ERR(tctx->io_wq);
7848 percpu_counter_destroy(&tctx->inflight);
7854 init_waitqueue_head(&tctx->wait);
7856 atomic_set(&tctx->in_idle, 0);
7857 task->io_uring = tctx;
7858 spin_lock_init(&tctx->task_lock);
7859 INIT_WQ_LIST(&tctx->task_list);
7860 tctx->task_state = 0;
7861 init_task_work(&tctx->task_work, tctx_task_work);
7865 void __io_uring_free(struct task_struct *tsk)
7867 struct io_uring_task *tctx = tsk->io_uring;
7869 WARN_ON_ONCE(!xa_empty(&tctx->xa));
7870 WARN_ON_ONCE(tctx->io_wq);
7872 percpu_counter_destroy(&tctx->inflight);
7874 tsk->io_uring = NULL;
7877 static int io_sq_offload_create(struct io_ring_ctx *ctx,
7878 struct io_uring_params *p)
7882 /* Retain compatibility with failing for an invalid attach attempt */
7883 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
7884 IORING_SETUP_ATTACH_WQ) {
7887 f = fdget(p->wq_fd);
7890 if (f.file->f_op != &io_uring_fops) {
7896 if (ctx->flags & IORING_SETUP_SQPOLL) {
7897 struct task_struct *tsk;
7898 struct io_sq_data *sqd;
7902 if (!capable(CAP_SYS_ADMIN) && !capable(CAP_SYS_NICE))
7905 sqd = io_get_sq_data(p, &attached);
7911 ctx->sq_creds = get_current_cred();
7913 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7914 if (!ctx->sq_thread_idle)
7915 ctx->sq_thread_idle = HZ;
7918 io_sq_thread_park(sqd);
7919 list_add(&ctx->sqd_list, &sqd->ctx_list);
7920 io_sqd_update_thread_idle(sqd);
7921 /* don't attach to a dying SQPOLL thread, would be racy */
7922 if (attached && !sqd->thread)
7924 io_sq_thread_unpark(sqd);
7931 if (p->flags & IORING_SETUP_SQ_AFF) {
7932 int cpu = p->sq_thread_cpu;
7935 if (cpu >= nr_cpu_ids)
7937 if (!cpu_online(cpu))
7945 sqd->task_pid = current->pid;
7946 sqd->task_tgid = current->tgid;
7947 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
7954 ret = io_uring_alloc_task_context(tsk, ctx);
7955 wake_up_new_task(tsk);
7958 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7959 /* Can't have SQ_AFF without SQPOLL */
7966 io_sq_thread_finish(ctx);
7969 complete(&ctx->sq_data->exited);
7973 static inline void __io_unaccount_mem(struct user_struct *user,
7974 unsigned long nr_pages)
7976 atomic_long_sub(nr_pages, &user->locked_vm);
7979 static inline int __io_account_mem(struct user_struct *user,
7980 unsigned long nr_pages)
7982 unsigned long page_limit, cur_pages, new_pages;
7984 /* Don't allow more pages than we can safely lock */
7985 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
7988 cur_pages = atomic_long_read(&user->locked_vm);
7989 new_pages = cur_pages + nr_pages;
7990 if (new_pages > page_limit)
7992 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7993 new_pages) != cur_pages);
7998 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8001 __io_unaccount_mem(ctx->user, nr_pages);
8003 if (ctx->mm_account)
8004 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8007 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8012 ret = __io_account_mem(ctx->user, nr_pages);
8017 if (ctx->mm_account)
8018 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8023 static void io_mem_free(void *ptr)
8030 page = virt_to_head_page(ptr);
8031 if (put_page_testzero(page))
8032 free_compound_page(page);
8035 static void *io_mem_alloc(size_t size)
8037 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8038 __GFP_NORETRY | __GFP_ACCOUNT;
8040 return (void *) __get_free_pages(gfp_flags, get_order(size));
8043 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8046 struct io_rings *rings;
8047 size_t off, sq_array_size;
8049 off = struct_size(rings, cqes, cq_entries);
8050 if (off == SIZE_MAX)
8054 off = ALIGN(off, SMP_CACHE_BYTES);
8062 sq_array_size = array_size(sizeof(u32), sq_entries);
8063 if (sq_array_size == SIZE_MAX)
8066 if (check_add_overflow(off, sq_array_size, &off))
8072 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8076 if (!ctx->user_bufs)
8079 for (i = 0; i < ctx->nr_user_bufs; i++) {
8080 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8082 for (j = 0; j < imu->nr_bvecs; j++)
8083 unpin_user_page(imu->bvec[j].bv_page);
8085 if (imu->acct_pages)
8086 io_unaccount_mem(ctx, imu->acct_pages);
8091 kfree(ctx->user_bufs);
8092 ctx->user_bufs = NULL;
8093 ctx->nr_user_bufs = 0;
8097 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8098 void __user *arg, unsigned index)
8100 struct iovec __user *src;
8102 #ifdef CONFIG_COMPAT
8104 struct compat_iovec __user *ciovs;
8105 struct compat_iovec ciov;
8107 ciovs = (struct compat_iovec __user *) arg;
8108 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8111 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8112 dst->iov_len = ciov.iov_len;
8116 src = (struct iovec __user *) arg;
8117 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8123 * Not super efficient, but this is just a registration time. And we do cache
8124 * the last compound head, so generally we'll only do a full search if we don't
8127 * We check if the given compound head page has already been accounted, to
8128 * avoid double accounting it. This allows us to account the full size of the
8129 * page, not just the constituent pages of a huge page.
8131 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8132 int nr_pages, struct page *hpage)
8136 /* check current page array */
8137 for (i = 0; i < nr_pages; i++) {
8138 if (!PageCompound(pages[i]))
8140 if (compound_head(pages[i]) == hpage)
8144 /* check previously registered pages */
8145 for (i = 0; i < ctx->nr_user_bufs; i++) {
8146 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8148 for (j = 0; j < imu->nr_bvecs; j++) {
8149 if (!PageCompound(imu->bvec[j].bv_page))
8151 if (compound_head(imu->bvec[j].bv_page) == hpage)
8159 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8160 int nr_pages, struct io_mapped_ubuf *imu,
8161 struct page **last_hpage)
8165 for (i = 0; i < nr_pages; i++) {
8166 if (!PageCompound(pages[i])) {
8171 hpage = compound_head(pages[i]);
8172 if (hpage == *last_hpage)
8174 *last_hpage = hpage;
8175 if (headpage_already_acct(ctx, pages, i, hpage))
8177 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8181 if (!imu->acct_pages)
8184 ret = io_account_mem(ctx, imu->acct_pages);
8186 imu->acct_pages = 0;
8190 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8191 struct io_mapped_ubuf *imu,
8192 struct page **last_hpage)
8194 struct vm_area_struct **vmas = NULL;
8195 struct page **pages = NULL;
8196 unsigned long off, start, end, ubuf;
8198 int ret, pret, nr_pages, i;
8200 ubuf = (unsigned long) iov->iov_base;
8201 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8202 start = ubuf >> PAGE_SHIFT;
8203 nr_pages = end - start;
8207 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8211 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8216 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
8222 mmap_read_lock(current->mm);
8223 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8225 if (pret == nr_pages) {
8226 /* don't support file backed memory */
8227 for (i = 0; i < nr_pages; i++) {
8228 struct vm_area_struct *vma = vmas[i];
8231 !is_file_hugepages(vma->vm_file)) {
8237 ret = pret < 0 ? pret : -EFAULT;
8239 mmap_read_unlock(current->mm);
8242 * if we did partial map, or found file backed vmas,
8243 * release any pages we did get
8246 unpin_user_pages(pages, pret);
8251 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8253 unpin_user_pages(pages, pret);
8258 off = ubuf & ~PAGE_MASK;
8259 size = iov->iov_len;
8260 for (i = 0; i < nr_pages; i++) {
8263 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8264 imu->bvec[i].bv_page = pages[i];
8265 imu->bvec[i].bv_len = vec_len;
8266 imu->bvec[i].bv_offset = off;
8270 /* store original address for later verification */
8272 imu->len = iov->iov_len;
8273 imu->nr_bvecs = nr_pages;
8281 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8285 if (!nr_args || nr_args > UIO_MAXIOV)
8288 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
8290 if (!ctx->user_bufs)
8296 static int io_buffer_validate(struct iovec *iov)
8299 * Don't impose further limits on the size and buffer
8300 * constraints here, we'll -EINVAL later when IO is
8301 * submitted if they are wrong.
8303 if (!iov->iov_base || !iov->iov_len)
8306 /* arbitrary limit, but we need something */
8307 if (iov->iov_len > SZ_1G)
8313 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8314 unsigned int nr_args)
8318 struct page *last_hpage = NULL;
8320 ret = io_buffers_map_alloc(ctx, nr_args);
8324 for (i = 0; i < nr_args; i++) {
8325 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8327 ret = io_copy_iov(ctx, &iov, arg, i);
8331 ret = io_buffer_validate(&iov);
8335 ret = io_sqe_buffer_register(ctx, &iov, imu, &last_hpage);
8339 ctx->nr_user_bufs++;
8343 io_sqe_buffers_unregister(ctx);
8348 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8350 __s32 __user *fds = arg;
8356 if (copy_from_user(&fd, fds, sizeof(*fds)))
8359 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8360 if (IS_ERR(ctx->cq_ev_fd)) {
8361 int ret = PTR_ERR(ctx->cq_ev_fd);
8362 ctx->cq_ev_fd = NULL;
8369 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8371 if (ctx->cq_ev_fd) {
8372 eventfd_ctx_put(ctx->cq_ev_fd);
8373 ctx->cq_ev_fd = NULL;
8380 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8382 struct io_buffer *buf;
8383 unsigned long index;
8385 xa_for_each(&ctx->io_buffers, index, buf)
8386 __io_remove_buffers(ctx, buf, index, -1U);
8389 static void io_req_cache_free(struct list_head *list, struct task_struct *tsk)
8391 struct io_kiocb *req, *nxt;
8393 list_for_each_entry_safe(req, nxt, list, compl.list) {
8394 if (tsk && req->task != tsk)
8396 list_del(&req->compl.list);
8397 kmem_cache_free(req_cachep, req);
8401 static void io_req_caches_free(struct io_ring_ctx *ctx)
8403 struct io_submit_state *submit_state = &ctx->submit_state;
8404 struct io_comp_state *cs = &ctx->submit_state.comp;
8406 mutex_lock(&ctx->uring_lock);
8408 if (submit_state->free_reqs) {
8409 kmem_cache_free_bulk(req_cachep, submit_state->free_reqs,
8410 submit_state->reqs);
8411 submit_state->free_reqs = 0;
8414 spin_lock_irq(&ctx->completion_lock);
8415 list_splice_init(&cs->locked_free_list, &cs->free_list);
8416 cs->locked_free_nr = 0;
8417 spin_unlock_irq(&ctx->completion_lock);
8419 io_req_cache_free(&cs->free_list, NULL);
8421 mutex_unlock(&ctx->uring_lock);
8424 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8427 * Some may use context even when all refs and requests have been put,
8428 * and they are free to do so while still holding uring_lock or
8429 * completion_lock, see __io_req_task_submit(). Wait for them to finish.
8431 mutex_lock(&ctx->uring_lock);
8432 mutex_unlock(&ctx->uring_lock);
8433 spin_lock_irq(&ctx->completion_lock);
8434 spin_unlock_irq(&ctx->completion_lock);
8436 io_sq_thread_finish(ctx);
8437 io_sqe_buffers_unregister(ctx);
8439 if (ctx->mm_account) {
8440 mmdrop(ctx->mm_account);
8441 ctx->mm_account = NULL;
8444 mutex_lock(&ctx->uring_lock);
8445 io_sqe_files_unregister(ctx);
8446 mutex_unlock(&ctx->uring_lock);
8447 io_eventfd_unregister(ctx);
8448 io_destroy_buffers(ctx);
8450 #if defined(CONFIG_UNIX)
8451 if (ctx->ring_sock) {
8452 ctx->ring_sock->file = NULL; /* so that iput() is called */
8453 sock_release(ctx->ring_sock);
8457 io_mem_free(ctx->rings);
8458 io_mem_free(ctx->sq_sqes);
8460 percpu_ref_exit(&ctx->refs);
8461 free_uid(ctx->user);
8462 io_req_caches_free(ctx);
8464 io_wq_put_hash(ctx->hash_map);
8465 kfree(ctx->cancel_hash);
8469 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8471 struct io_ring_ctx *ctx = file->private_data;
8474 poll_wait(file, &ctx->cq_wait, wait);
8476 * synchronizes with barrier from wq_has_sleeper call in
8480 if (!io_sqring_full(ctx))
8481 mask |= EPOLLOUT | EPOLLWRNORM;
8484 * Don't flush cqring overflow list here, just do a simple check.
8485 * Otherwise there could possible be ABBA deadlock:
8488 * lock(&ctx->uring_lock);
8490 * lock(&ctx->uring_lock);
8493 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8494 * pushs them to do the flush.
8496 if (io_cqring_events(ctx) || test_bit(0, &ctx->cq_check_overflow))
8497 mask |= EPOLLIN | EPOLLRDNORM;
8502 static int io_uring_fasync(int fd, struct file *file, int on)
8504 struct io_ring_ctx *ctx = file->private_data;
8506 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8509 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8511 const struct cred *creds;
8513 creds = xa_erase(&ctx->personalities, id);
8522 static inline bool io_run_ctx_fallback(struct io_ring_ctx *ctx)
8524 return io_run_task_work_head(&ctx->exit_task_work);
8527 struct io_tctx_exit {
8528 struct callback_head task_work;
8529 struct completion completion;
8530 struct io_ring_ctx *ctx;
8533 static void io_tctx_exit_cb(struct callback_head *cb)
8535 struct io_uring_task *tctx = current->io_uring;
8536 struct io_tctx_exit *work;
8538 work = container_of(cb, struct io_tctx_exit, task_work);
8540 * When @in_idle, we're in cancellation and it's racy to remove the
8541 * node. It'll be removed by the end of cancellation, just ignore it.
8543 if (!atomic_read(&tctx->in_idle))
8544 io_uring_del_task_file((unsigned long)work->ctx);
8545 complete(&work->completion);
8548 static void io_ring_exit_work(struct work_struct *work)
8550 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
8551 unsigned long timeout = jiffies + HZ * 60 * 5;
8552 struct io_tctx_exit exit;
8553 struct io_tctx_node *node;
8556 /* prevent SQPOLL from submitting new requests */
8558 io_sq_thread_park(ctx->sq_data);
8559 list_del_init(&ctx->sqd_list);
8560 io_sqd_update_thread_idle(ctx->sq_data);
8561 io_sq_thread_unpark(ctx->sq_data);
8565 * If we're doing polled IO and end up having requests being
8566 * submitted async (out-of-line), then completions can come in while
8567 * we're waiting for refs to drop. We need to reap these manually,
8568 * as nobody else will be looking for them.
8571 io_uring_try_cancel_requests(ctx, NULL, NULL);
8573 WARN_ON_ONCE(time_after(jiffies, timeout));
8574 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8576 mutex_lock(&ctx->uring_lock);
8577 while (!list_empty(&ctx->tctx_list)) {
8578 WARN_ON_ONCE(time_after(jiffies, timeout));
8580 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
8583 init_completion(&exit.completion);
8584 init_task_work(&exit.task_work, io_tctx_exit_cb);
8585 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
8586 if (WARN_ON_ONCE(ret))
8588 wake_up_process(node->task);
8590 mutex_unlock(&ctx->uring_lock);
8591 wait_for_completion(&exit.completion);
8593 mutex_lock(&ctx->uring_lock);
8595 mutex_unlock(&ctx->uring_lock);
8597 io_ring_ctx_free(ctx);
8600 /* Returns true if we found and killed one or more timeouts */
8601 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
8602 struct files_struct *files)
8604 struct io_kiocb *req, *tmp;
8607 spin_lock_irq(&ctx->completion_lock);
8608 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
8609 if (io_match_task(req, tsk, files)) {
8610 io_kill_timeout(req, -ECANCELED);
8615 io_commit_cqring(ctx);
8616 spin_unlock_irq(&ctx->completion_lock);
8618 io_cqring_ev_posted(ctx);
8619 return canceled != 0;
8622 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8624 unsigned long index;
8625 struct creds *creds;
8627 mutex_lock(&ctx->uring_lock);
8628 percpu_ref_kill(&ctx->refs);
8629 /* if force is set, the ring is going away. always drop after that */
8630 ctx->cq_overflow_flushed = 1;
8632 __io_cqring_overflow_flush(ctx, true, NULL, NULL);
8633 xa_for_each(&ctx->personalities, index, creds)
8634 io_unregister_personality(ctx, index);
8635 mutex_unlock(&ctx->uring_lock);
8637 io_kill_timeouts(ctx, NULL, NULL);
8638 io_poll_remove_all(ctx, NULL, NULL);
8640 /* if we failed setting up the ctx, we might not have any rings */
8641 io_iopoll_try_reap_events(ctx);
8643 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8645 * Use system_unbound_wq to avoid spawning tons of event kworkers
8646 * if we're exiting a ton of rings at the same time. It just adds
8647 * noise and overhead, there's no discernable change in runtime
8648 * over using system_wq.
8650 queue_work(system_unbound_wq, &ctx->exit_work);
8653 static int io_uring_release(struct inode *inode, struct file *file)
8655 struct io_ring_ctx *ctx = file->private_data;
8657 file->private_data = NULL;
8658 io_ring_ctx_wait_and_kill(ctx);
8662 struct io_task_cancel {
8663 struct task_struct *task;
8664 struct files_struct *files;
8667 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8669 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8670 struct io_task_cancel *cancel = data;
8673 if (cancel->files && (req->flags & REQ_F_LINK_TIMEOUT)) {
8674 unsigned long flags;
8675 struct io_ring_ctx *ctx = req->ctx;
8677 /* protect against races with linked timeouts */
8678 spin_lock_irqsave(&ctx->completion_lock, flags);
8679 ret = io_match_task(req, cancel->task, cancel->files);
8680 spin_unlock_irqrestore(&ctx->completion_lock, flags);
8682 ret = io_match_task(req, cancel->task, cancel->files);
8687 static bool io_cancel_defer_files(struct io_ring_ctx *ctx,
8688 struct task_struct *task,
8689 struct files_struct *files)
8691 struct io_defer_entry *de;
8694 spin_lock_irq(&ctx->completion_lock);
8695 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8696 if (io_match_task(de->req, task, files)) {
8697 list_cut_position(&list, &ctx->defer_list, &de->list);
8701 spin_unlock_irq(&ctx->completion_lock);
8702 if (list_empty(&list))
8705 while (!list_empty(&list)) {
8706 de = list_first_entry(&list, struct io_defer_entry, list);
8707 list_del_init(&de->list);
8708 req_set_fail_links(de->req);
8709 io_put_req(de->req);
8710 io_req_complete(de->req, -ECANCELED);
8716 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
8718 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8720 return req->ctx == data;
8723 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
8725 struct io_tctx_node *node;
8726 enum io_wq_cancel cret;
8729 mutex_lock(&ctx->uring_lock);
8730 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
8731 struct io_uring_task *tctx = node->task->io_uring;
8734 * io_wq will stay alive while we hold uring_lock, because it's
8735 * killed after ctx nodes, which requires to take the lock.
8737 if (!tctx || !tctx->io_wq)
8739 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
8740 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8742 mutex_unlock(&ctx->uring_lock);
8747 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
8748 struct task_struct *task,
8749 struct files_struct *files)
8751 struct io_task_cancel cancel = { .task = task, .files = files, };
8752 struct io_uring_task *tctx = task ? task->io_uring : NULL;
8755 enum io_wq_cancel cret;
8759 ret |= io_uring_try_cancel_iowq(ctx);
8760 } else if (tctx && tctx->io_wq) {
8762 * Cancels requests of all rings, not only @ctx, but
8763 * it's fine as the task is in exit/exec.
8765 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
8767 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8770 /* SQPOLL thread does its own polling */
8771 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && !files) ||
8772 (ctx->sq_data && ctx->sq_data->thread == current)) {
8773 while (!list_empty_careful(&ctx->iopoll_list)) {
8774 io_iopoll_try_reap_events(ctx);
8779 ret |= io_cancel_defer_files(ctx, task, files);
8780 ret |= io_poll_remove_all(ctx, task, files);
8781 ret |= io_kill_timeouts(ctx, task, files);
8782 ret |= io_run_task_work();
8783 ret |= io_run_ctx_fallback(ctx);
8784 io_cqring_overflow_flush(ctx, true, task, files);
8791 static int io_uring_count_inflight(struct io_ring_ctx *ctx,
8792 struct task_struct *task,
8793 struct files_struct *files)
8795 struct io_kiocb *req;
8798 spin_lock_irq(&ctx->inflight_lock);
8799 list_for_each_entry(req, &ctx->inflight_list, inflight_entry)
8800 cnt += io_match_task(req, task, files);
8801 spin_unlock_irq(&ctx->inflight_lock);
8805 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
8806 struct task_struct *task,
8807 struct files_struct *files)
8809 while (!list_empty_careful(&ctx->inflight_list)) {
8813 inflight = io_uring_count_inflight(ctx, task, files);
8817 io_uring_try_cancel_requests(ctx, task, files);
8819 prepare_to_wait(&task->io_uring->wait, &wait,
8820 TASK_UNINTERRUPTIBLE);
8821 if (inflight == io_uring_count_inflight(ctx, task, files))
8823 finish_wait(&task->io_uring->wait, &wait);
8828 * Note that this task has used io_uring. We use it for cancelation purposes.
8830 static int io_uring_add_task_file(struct io_ring_ctx *ctx)
8832 struct io_uring_task *tctx = current->io_uring;
8833 struct io_tctx_node *node;
8836 if (unlikely(!tctx)) {
8837 ret = io_uring_alloc_task_context(current, ctx);
8840 tctx = current->io_uring;
8842 if (tctx->last != ctx) {
8843 void *old = xa_load(&tctx->xa, (unsigned long)ctx);
8846 node = kmalloc(sizeof(*node), GFP_KERNEL);
8850 node->task = current;
8852 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
8859 mutex_lock(&ctx->uring_lock);
8860 list_add(&node->ctx_node, &ctx->tctx_list);
8861 mutex_unlock(&ctx->uring_lock);
8869 * Remove this io_uring_file -> task mapping.
8871 static void io_uring_del_task_file(unsigned long index)
8873 struct io_uring_task *tctx = current->io_uring;
8874 struct io_tctx_node *node;
8878 node = xa_erase(&tctx->xa, index);
8882 WARN_ON_ONCE(current != node->task);
8883 WARN_ON_ONCE(list_empty(&node->ctx_node));
8885 mutex_lock(&node->ctx->uring_lock);
8886 list_del(&node->ctx_node);
8887 mutex_unlock(&node->ctx->uring_lock);
8889 if (tctx->last == node->ctx)
8894 static void io_uring_clean_tctx(struct io_uring_task *tctx)
8896 struct io_tctx_node *node;
8897 unsigned long index;
8899 xa_for_each(&tctx->xa, index, node)
8900 io_uring_del_task_file(index);
8902 io_wq_put_and_exit(tctx->io_wq);
8907 static s64 tctx_inflight(struct io_uring_task *tctx)
8909 return percpu_counter_sum(&tctx->inflight);
8912 static void io_sqpoll_cancel_cb(struct callback_head *cb)
8914 struct io_tctx_exit *work = container_of(cb, struct io_tctx_exit, task_work);
8915 struct io_ring_ctx *ctx = work->ctx;
8916 struct io_sq_data *sqd = ctx->sq_data;
8919 io_uring_cancel_sqpoll(ctx);
8920 complete(&work->completion);
8923 static void io_sqpoll_cancel_sync(struct io_ring_ctx *ctx)
8925 struct io_sq_data *sqd = ctx->sq_data;
8926 struct io_tctx_exit work = { .ctx = ctx, };
8927 struct task_struct *task;
8929 io_sq_thread_park(sqd);
8930 list_del_init(&ctx->sqd_list);
8931 io_sqd_update_thread_idle(sqd);
8934 init_completion(&work.completion);
8935 init_task_work(&work.task_work, io_sqpoll_cancel_cb);
8936 io_task_work_add_head(&sqd->park_task_work, &work.task_work);
8937 wake_up_process(task);
8939 io_sq_thread_unpark(sqd);
8942 wait_for_completion(&work.completion);
8945 void __io_uring_files_cancel(struct files_struct *files)
8947 struct io_uring_task *tctx = current->io_uring;
8948 struct io_tctx_node *node;
8949 unsigned long index;
8951 /* make sure overflow events are dropped */
8952 atomic_inc(&tctx->in_idle);
8953 xa_for_each(&tctx->xa, index, node) {
8954 struct io_ring_ctx *ctx = node->ctx;
8957 io_sqpoll_cancel_sync(ctx);
8960 io_uring_cancel_files(ctx, current, files);
8962 io_uring_try_cancel_requests(ctx, current, NULL);
8964 atomic_dec(&tctx->in_idle);
8967 io_uring_clean_tctx(tctx);
8970 /* should only be called by SQPOLL task */
8971 static void io_uring_cancel_sqpoll(struct io_ring_ctx *ctx)
8973 struct io_sq_data *sqd = ctx->sq_data;
8974 struct io_uring_task *tctx = current->io_uring;
8978 WARN_ON_ONCE(!sqd || ctx->sq_data->thread != current);
8980 atomic_inc(&tctx->in_idle);
8982 /* read completions before cancelations */
8983 inflight = tctx_inflight(tctx);
8986 io_uring_try_cancel_requests(ctx, current, NULL);
8988 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8990 * If we've seen completions, retry without waiting. This
8991 * avoids a race where a completion comes in before we did
8992 * prepare_to_wait().
8994 if (inflight == tctx_inflight(tctx))
8996 finish_wait(&tctx->wait, &wait);
8998 atomic_dec(&tctx->in_idle);
9002 * Find any io_uring fd that this task has registered or done IO on, and cancel
9005 void __io_uring_task_cancel(void)
9007 struct io_uring_task *tctx = current->io_uring;
9011 /* make sure overflow events are dropped */
9012 atomic_inc(&tctx->in_idle);
9013 __io_uring_files_cancel(NULL);
9016 /* read completions before cancelations */
9017 inflight = tctx_inflight(tctx);
9020 __io_uring_files_cancel(NULL);
9022 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9025 * If we've seen completions, retry without waiting. This
9026 * avoids a race where a completion comes in before we did
9027 * prepare_to_wait().
9029 if (inflight == tctx_inflight(tctx))
9031 finish_wait(&tctx->wait, &wait);
9034 atomic_dec(&tctx->in_idle);
9036 io_uring_clean_tctx(tctx);
9037 /* all current's requests should be gone, we can kill tctx */
9038 __io_uring_free(current);
9041 static void *io_uring_validate_mmap_request(struct file *file,
9042 loff_t pgoff, size_t sz)
9044 struct io_ring_ctx *ctx = file->private_data;
9045 loff_t offset = pgoff << PAGE_SHIFT;
9050 case IORING_OFF_SQ_RING:
9051 case IORING_OFF_CQ_RING:
9054 case IORING_OFF_SQES:
9058 return ERR_PTR(-EINVAL);
9061 page = virt_to_head_page(ptr);
9062 if (sz > page_size(page))
9063 return ERR_PTR(-EINVAL);
9070 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9072 size_t sz = vma->vm_end - vma->vm_start;
9076 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9078 return PTR_ERR(ptr);
9080 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9081 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9084 #else /* !CONFIG_MMU */
9086 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9088 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9091 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9093 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9096 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9097 unsigned long addr, unsigned long len,
9098 unsigned long pgoff, unsigned long flags)
9102 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9104 return PTR_ERR(ptr);
9106 return (unsigned long) ptr;
9109 #endif /* !CONFIG_MMU */
9111 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9116 if (!io_sqring_full(ctx))
9118 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9120 if (!io_sqring_full(ctx))
9123 } while (!signal_pending(current));
9125 finish_wait(&ctx->sqo_sq_wait, &wait);
9129 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9130 struct __kernel_timespec __user **ts,
9131 const sigset_t __user **sig)
9133 struct io_uring_getevents_arg arg;
9136 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9137 * is just a pointer to the sigset_t.
9139 if (!(flags & IORING_ENTER_EXT_ARG)) {
9140 *sig = (const sigset_t __user *) argp;
9146 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9147 * timespec and sigset_t pointers if good.
9149 if (*argsz != sizeof(arg))
9151 if (copy_from_user(&arg, argp, sizeof(arg)))
9153 *sig = u64_to_user_ptr(arg.sigmask);
9154 *argsz = arg.sigmask_sz;
9155 *ts = u64_to_user_ptr(arg.ts);
9159 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9160 u32, min_complete, u32, flags, const void __user *, argp,
9163 struct io_ring_ctx *ctx;
9170 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9171 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG))
9179 if (f.file->f_op != &io_uring_fops)
9183 ctx = f.file->private_data;
9184 if (!percpu_ref_tryget(&ctx->refs))
9188 if (ctx->flags & IORING_SETUP_R_DISABLED)
9192 * For SQ polling, the thread will do all submissions and completions.
9193 * Just return the requested submit count, and wake the thread if
9197 if (ctx->flags & IORING_SETUP_SQPOLL) {
9198 io_cqring_overflow_flush(ctx, false, NULL, NULL);
9201 if (unlikely(ctx->sq_data->thread == NULL)) {
9204 if (flags & IORING_ENTER_SQ_WAKEUP)
9205 wake_up(&ctx->sq_data->wait);
9206 if (flags & IORING_ENTER_SQ_WAIT) {
9207 ret = io_sqpoll_wait_sq(ctx);
9211 submitted = to_submit;
9212 } else if (to_submit) {
9213 ret = io_uring_add_task_file(ctx);
9216 mutex_lock(&ctx->uring_lock);
9217 submitted = io_submit_sqes(ctx, to_submit);
9218 mutex_unlock(&ctx->uring_lock);
9220 if (submitted != to_submit)
9223 if (flags & IORING_ENTER_GETEVENTS) {
9224 const sigset_t __user *sig;
9225 struct __kernel_timespec __user *ts;
9227 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9231 min_complete = min(min_complete, ctx->cq_entries);
9234 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9235 * space applications don't need to do io completion events
9236 * polling again, they can rely on io_sq_thread to do polling
9237 * work, which can reduce cpu usage and uring_lock contention.
9239 if (ctx->flags & IORING_SETUP_IOPOLL &&
9240 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9241 ret = io_iopoll_check(ctx, min_complete);
9243 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9248 percpu_ref_put(&ctx->refs);
9251 return submitted ? submitted : ret;
9254 #ifdef CONFIG_PROC_FS
9255 static int io_uring_show_cred(struct seq_file *m, unsigned int id,
9256 const struct cred *cred)
9258 struct user_namespace *uns = seq_user_ns(m);
9259 struct group_info *gi;
9264 seq_printf(m, "%5d\n", id);
9265 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9266 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9267 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9268 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9269 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9270 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9271 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9272 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9273 seq_puts(m, "\n\tGroups:\t");
9274 gi = cred->group_info;
9275 for (g = 0; g < gi->ngroups; g++) {
9276 seq_put_decimal_ull(m, g ? " " : "",
9277 from_kgid_munged(uns, gi->gid[g]));
9279 seq_puts(m, "\n\tCapEff:\t");
9280 cap = cred->cap_effective;
9281 CAP_FOR_EACH_U32(__capi)
9282 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9287 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9289 struct io_sq_data *sq = NULL;
9294 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9295 * since fdinfo case grabs it in the opposite direction of normal use
9296 * cases. If we fail to get the lock, we just don't iterate any
9297 * structures that could be going away outside the io_uring mutex.
9299 has_lock = mutex_trylock(&ctx->uring_lock);
9301 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
9307 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9308 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9309 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9310 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9311 struct file *f = *io_fixed_file_slot(ctx->file_data, i);
9314 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9316 seq_printf(m, "%5u: <none>\n", i);
9318 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9319 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9320 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
9322 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
9323 (unsigned int) buf->len);
9325 if (has_lock && !xa_empty(&ctx->personalities)) {
9326 unsigned long index;
9327 const struct cred *cred;
9329 seq_printf(m, "Personalities:\n");
9330 xa_for_each(&ctx->personalities, index, cred)
9331 io_uring_show_cred(m, index, cred);
9333 seq_printf(m, "PollList:\n");
9334 spin_lock_irq(&ctx->completion_lock);
9335 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9336 struct hlist_head *list = &ctx->cancel_hash[i];
9337 struct io_kiocb *req;
9339 hlist_for_each_entry(req, list, hash_node)
9340 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9341 req->task->task_works != NULL);
9343 spin_unlock_irq(&ctx->completion_lock);
9345 mutex_unlock(&ctx->uring_lock);
9348 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9350 struct io_ring_ctx *ctx = f->private_data;
9352 if (percpu_ref_tryget(&ctx->refs)) {
9353 __io_uring_show_fdinfo(ctx, m);
9354 percpu_ref_put(&ctx->refs);
9359 static const struct file_operations io_uring_fops = {
9360 .release = io_uring_release,
9361 .mmap = io_uring_mmap,
9363 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9364 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9366 .poll = io_uring_poll,
9367 .fasync = io_uring_fasync,
9368 #ifdef CONFIG_PROC_FS
9369 .show_fdinfo = io_uring_show_fdinfo,
9373 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9374 struct io_uring_params *p)
9376 struct io_rings *rings;
9377 size_t size, sq_array_offset;
9379 /* make sure these are sane, as we already accounted them */
9380 ctx->sq_entries = p->sq_entries;
9381 ctx->cq_entries = p->cq_entries;
9383 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9384 if (size == SIZE_MAX)
9387 rings = io_mem_alloc(size);
9392 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9393 rings->sq_ring_mask = p->sq_entries - 1;
9394 rings->cq_ring_mask = p->cq_entries - 1;
9395 rings->sq_ring_entries = p->sq_entries;
9396 rings->cq_ring_entries = p->cq_entries;
9397 ctx->sq_mask = rings->sq_ring_mask;
9398 ctx->cq_mask = rings->cq_ring_mask;
9400 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9401 if (size == SIZE_MAX) {
9402 io_mem_free(ctx->rings);
9407 ctx->sq_sqes = io_mem_alloc(size);
9408 if (!ctx->sq_sqes) {
9409 io_mem_free(ctx->rings);
9417 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9421 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9425 ret = io_uring_add_task_file(ctx);
9430 fd_install(fd, file);
9435 * Allocate an anonymous fd, this is what constitutes the application
9436 * visible backing of an io_uring instance. The application mmaps this
9437 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9438 * we have to tie this fd to a socket for file garbage collection purposes.
9440 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
9443 #if defined(CONFIG_UNIX)
9446 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9449 return ERR_PTR(ret);
9452 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9453 O_RDWR | O_CLOEXEC);
9454 #if defined(CONFIG_UNIX)
9456 sock_release(ctx->ring_sock);
9457 ctx->ring_sock = NULL;
9459 ctx->ring_sock->file = file;
9465 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9466 struct io_uring_params __user *params)
9468 struct io_ring_ctx *ctx;
9474 if (entries > IORING_MAX_ENTRIES) {
9475 if (!(p->flags & IORING_SETUP_CLAMP))
9477 entries = IORING_MAX_ENTRIES;
9481 * Use twice as many entries for the CQ ring. It's possible for the
9482 * application to drive a higher depth than the size of the SQ ring,
9483 * since the sqes are only used at submission time. This allows for
9484 * some flexibility in overcommitting a bit. If the application has
9485 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9486 * of CQ ring entries manually.
9488 p->sq_entries = roundup_pow_of_two(entries);
9489 if (p->flags & IORING_SETUP_CQSIZE) {
9491 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9492 * to a power-of-two, if it isn't already. We do NOT impose
9493 * any cq vs sq ring sizing.
9497 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9498 if (!(p->flags & IORING_SETUP_CLAMP))
9500 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9502 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9503 if (p->cq_entries < p->sq_entries)
9506 p->cq_entries = 2 * p->sq_entries;
9509 ctx = io_ring_ctx_alloc(p);
9512 ctx->compat = in_compat_syscall();
9513 if (!capable(CAP_IPC_LOCK))
9514 ctx->user = get_uid(current_user());
9517 * This is just grabbed for accounting purposes. When a process exits,
9518 * the mm is exited and dropped before the files, hence we need to hang
9519 * on to this mm purely for the purposes of being able to unaccount
9520 * memory (locked/pinned vm). It's not used for anything else.
9522 mmgrab(current->mm);
9523 ctx->mm_account = current->mm;
9525 ret = io_allocate_scq_urings(ctx, p);
9529 ret = io_sq_offload_create(ctx, p);
9533 memset(&p->sq_off, 0, sizeof(p->sq_off));
9534 p->sq_off.head = offsetof(struct io_rings, sq.head);
9535 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9536 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9537 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9538 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9539 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9540 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9542 memset(&p->cq_off, 0, sizeof(p->cq_off));
9543 p->cq_off.head = offsetof(struct io_rings, cq.head);
9544 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9545 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9546 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9547 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9548 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9549 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9551 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9552 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9553 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9554 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9555 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS;
9557 if (copy_to_user(params, p, sizeof(*p))) {
9562 file = io_uring_get_file(ctx);
9564 ret = PTR_ERR(file);
9569 * Install ring fd as the very last thing, so we don't risk someone
9570 * having closed it before we finish setup
9572 ret = io_uring_install_fd(ctx, file);
9574 /* fput will clean it up */
9579 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9582 io_ring_ctx_wait_and_kill(ctx);
9587 * Sets up an aio uring context, and returns the fd. Applications asks for a
9588 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9589 * params structure passed in.
9591 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9593 struct io_uring_params p;
9596 if (copy_from_user(&p, params, sizeof(p)))
9598 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9603 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9604 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9605 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9606 IORING_SETUP_R_DISABLED))
9609 return io_uring_create(entries, &p, params);
9612 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9613 struct io_uring_params __user *, params)
9615 return io_uring_setup(entries, params);
9618 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9620 struct io_uring_probe *p;
9624 size = struct_size(p, ops, nr_args);
9625 if (size == SIZE_MAX)
9627 p = kzalloc(size, GFP_KERNEL);
9632 if (copy_from_user(p, arg, size))
9635 if (memchr_inv(p, 0, size))
9638 p->last_op = IORING_OP_LAST - 1;
9639 if (nr_args > IORING_OP_LAST)
9640 nr_args = IORING_OP_LAST;
9642 for (i = 0; i < nr_args; i++) {
9644 if (!io_op_defs[i].not_supported)
9645 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9650 if (copy_to_user(arg, p, size))
9657 static int io_register_personality(struct io_ring_ctx *ctx)
9659 const struct cred *creds;
9663 creds = get_current_cred();
9665 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
9666 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
9673 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9674 unsigned int nr_args)
9676 struct io_uring_restriction *res;
9680 /* Restrictions allowed only if rings started disabled */
9681 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9684 /* We allow only a single restrictions registration */
9685 if (ctx->restrictions.registered)
9688 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9691 size = array_size(nr_args, sizeof(*res));
9692 if (size == SIZE_MAX)
9695 res = memdup_user(arg, size);
9697 return PTR_ERR(res);
9701 for (i = 0; i < nr_args; i++) {
9702 switch (res[i].opcode) {
9703 case IORING_RESTRICTION_REGISTER_OP:
9704 if (res[i].register_op >= IORING_REGISTER_LAST) {
9709 __set_bit(res[i].register_op,
9710 ctx->restrictions.register_op);
9712 case IORING_RESTRICTION_SQE_OP:
9713 if (res[i].sqe_op >= IORING_OP_LAST) {
9718 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9720 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9721 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9723 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9724 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9733 /* Reset all restrictions if an error happened */
9735 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9737 ctx->restrictions.registered = true;
9743 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9745 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9748 if (ctx->restrictions.registered)
9749 ctx->restricted = 1;
9751 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9752 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
9753 wake_up(&ctx->sq_data->wait);
9757 static bool io_register_op_must_quiesce(int op)
9760 case IORING_UNREGISTER_FILES:
9761 case IORING_REGISTER_FILES_UPDATE:
9762 case IORING_REGISTER_PROBE:
9763 case IORING_REGISTER_PERSONALITY:
9764 case IORING_UNREGISTER_PERSONALITY:
9771 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
9772 void __user *arg, unsigned nr_args)
9773 __releases(ctx->uring_lock)
9774 __acquires(ctx->uring_lock)
9779 * We're inside the ring mutex, if the ref is already dying, then
9780 * someone else killed the ctx or is already going through
9781 * io_uring_register().
9783 if (percpu_ref_is_dying(&ctx->refs))
9786 if (io_register_op_must_quiesce(opcode)) {
9787 percpu_ref_kill(&ctx->refs);
9790 * Drop uring mutex before waiting for references to exit. If
9791 * another thread is currently inside io_uring_enter() it might
9792 * need to grab the uring_lock to make progress. If we hold it
9793 * here across the drain wait, then we can deadlock. It's safe
9794 * to drop the mutex here, since no new references will come in
9795 * after we've killed the percpu ref.
9797 mutex_unlock(&ctx->uring_lock);
9799 ret = wait_for_completion_interruptible(&ctx->ref_comp);
9802 ret = io_run_task_work_sig();
9807 mutex_lock(&ctx->uring_lock);
9810 percpu_ref_resurrect(&ctx->refs);
9815 if (ctx->restricted) {
9816 if (opcode >= IORING_REGISTER_LAST) {
9821 if (!test_bit(opcode, ctx->restrictions.register_op)) {
9828 case IORING_REGISTER_BUFFERS:
9829 ret = io_sqe_buffers_register(ctx, arg, nr_args);
9831 case IORING_UNREGISTER_BUFFERS:
9835 ret = io_sqe_buffers_unregister(ctx);
9837 case IORING_REGISTER_FILES:
9838 ret = io_sqe_files_register(ctx, arg, nr_args);
9840 case IORING_UNREGISTER_FILES:
9844 ret = io_sqe_files_unregister(ctx);
9846 case IORING_REGISTER_FILES_UPDATE:
9847 ret = io_sqe_files_update(ctx, arg, nr_args);
9849 case IORING_REGISTER_EVENTFD:
9850 case IORING_REGISTER_EVENTFD_ASYNC:
9854 ret = io_eventfd_register(ctx, arg);
9857 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
9858 ctx->eventfd_async = 1;
9860 ctx->eventfd_async = 0;
9862 case IORING_UNREGISTER_EVENTFD:
9866 ret = io_eventfd_unregister(ctx);
9868 case IORING_REGISTER_PROBE:
9870 if (!arg || nr_args > 256)
9872 ret = io_probe(ctx, arg, nr_args);
9874 case IORING_REGISTER_PERSONALITY:
9878 ret = io_register_personality(ctx);
9880 case IORING_UNREGISTER_PERSONALITY:
9884 ret = io_unregister_personality(ctx, nr_args);
9886 case IORING_REGISTER_ENABLE_RINGS:
9890 ret = io_register_enable_rings(ctx);
9892 case IORING_REGISTER_RESTRICTIONS:
9893 ret = io_register_restrictions(ctx, arg, nr_args);
9901 if (io_register_op_must_quiesce(opcode)) {
9902 /* bring the ctx back to life */
9903 percpu_ref_reinit(&ctx->refs);
9905 reinit_completion(&ctx->ref_comp);
9910 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
9911 void __user *, arg, unsigned int, nr_args)
9913 struct io_ring_ctx *ctx;
9922 if (f.file->f_op != &io_uring_fops)
9925 ctx = f.file->private_data;
9929 mutex_lock(&ctx->uring_lock);
9930 ret = __io_uring_register(ctx, opcode, arg, nr_args);
9931 mutex_unlock(&ctx->uring_lock);
9932 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
9933 ctx->cq_ev_fd != NULL, ret);
9939 static int __init io_uring_init(void)
9941 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
9942 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
9943 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
9946 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
9947 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
9948 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
9949 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
9950 BUILD_BUG_SQE_ELEM(1, __u8, flags);
9951 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
9952 BUILD_BUG_SQE_ELEM(4, __s32, fd);
9953 BUILD_BUG_SQE_ELEM(8, __u64, off);
9954 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
9955 BUILD_BUG_SQE_ELEM(16, __u64, addr);
9956 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
9957 BUILD_BUG_SQE_ELEM(24, __u32, len);
9958 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
9959 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
9960 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
9961 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
9962 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
9963 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
9964 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
9965 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
9966 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
9967 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
9968 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
9969 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
9970 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
9971 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
9972 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
9973 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
9974 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
9975 BUILD_BUG_SQE_ELEM(42, __u16, personality);
9976 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
9978 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
9979 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
9980 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
9984 __initcall(io_uring_init);